{"schema":"5lc-index-ledger/1","key":"models-open-weight","layer":"models","sublayer":"open-weight","asOf":"2026-10-22","about":"5LC Importance Index ledger for one sublayer: the measure card, the current ten with any editorial overrides and their reasons, the computed result for every scored company (Importance = share of AI output × years to replace, Reach, replacement years, Rising, the Monte Carlo rank simulation and its stability class, validation, the revenue baseline and rater agreement), each scored company's sourced inputs and irreplaceability ratings, and the sublayer's validation cases with their sources. Monte Carlo frequencies describe the stated input ranges, not true probabilities. Omitted: the free-text notes of the internal check passes and of rating reconciliations, and the location of the internal card review; references to local evidence files and to internal method decisions are replaced by plain descriptions.","roster":{"members":[{"entityId":"alibaba-qwen","slot":"importance","since":"2026-10-22"},{"entityId":"deepseek","slot":"importance","since":"2026-10-22"},{"entityId":"google-deepmind","slot":"importance","since":"2026-10-22"},{"entityId":"zhipu","slot":"importance","since":"2026-10-22"},{"entityId":"meta-superintelligence-labs","slot":"importance","since":"2026-10-22"},{"entityId":"nvidia-nemotron","slot":"importance","since":"2026-10-22"},{"entityId":"tencent-hy","slot":"importance","since":"2026-10-22"},{"entityId":"xiaomi","slot":"importance","since":"2026-10-22"},{"entityId":"openai","slot":"importance","since":"2026-10-22"},{"entityId":"mistral","slot":"importance","since":"2026-10-22"}],"pending":{},"lastRun":"2026-10-22","resultsDate":"2026-10-22"},"card":{"schema":"5lc-index-card/1","key":"models-open-weight","layer":"models","sublayer":"open-weight","version":1,"scope":"Use of open-weight language models (weights published for download under any licence), including code, reasoning and vision-language models, counted for the developer that trained the base weights; fine-tuned, quantised and merged copies count for the base model's developer, a developer's own API sales of its open models count here as hosted use, closed models, including regional ones, count in Frontier labs, serving revenue counts in Inference serving, open image, video and speech models count in Specialist models through their developers' revenue, and embedding models are not scored.","outputMeasure":"How much open-weight language models are used: tokens served through APIs, downloads for running on companies' own servers, and copies run on local devices.","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","hardToReplace":"Published weights stay available if their developer stops, and users can move to another open model within weeks. What is hard to replace is a steady supply of new, capable releases and the tools and fine-tunes built around one model family.","functions":[{"id":"hosted-inference","name":"Hosted use through APIs","why":"Open models bought per token from their developers or from inference providers.","weight":{"low":0.15,"central":0.261,"high":0.5},"weightSource":{"url":"https://www.linuxfoundation.org/hubfs/LF%20Research/lfr_genai24_111924.pdf","title":"Shaping the Future of Generative AI: The Impact of Open Source Innovation","publisher":"The Linux Foundation","publishedDate":"2024-11-19","accessedDate":"2026-10-22","tier":"T4","locator":"Figure 13, p. 24 (survey Aug–Sep 2024): organisations whose generative-AI infrastructure is more than 50% open source host inference through managed API services 46%, self-host in the cloud 56%, on premises 50%, on devices 24%"},"shareData":"Token shares from OpenRouter, Vercel AI Gateway and the ATOM Report; IDC China token volumes re-attributed to each model's base developer (platform figures are off-basis)"},{"id":"self-hosted","name":"Self-hosted in companies' own cloud or data centres","why":"Companies run downloaded weights on servers they control; the most common use.","weight":{"low":0.35,"central":0.602,"high":0.7},"weightSource":{"url":"https://www.linuxfoundation.org/hubfs/LF%20Research/lfr_genai24_111924.pdf","title":"Shaping the Future of Generative AI: The Impact of Open Source Innovation","publisher":"The Linux Foundation","publishedDate":"2024-11-19","accessedDate":"2026-10-22","tier":"T4","locator":"Figure 13, p. 24: self-host in the cloud 56% and on premises 50% among organisations with more than 50% open-source generative-AI infrastructure; Figure 12, p. 23: 49% and 38% of all respondents"},"shareData":"Download shares of full weights (Hugging Face, ModelScope; ATOM Report), checked against derivative-model shares"},{"id":"local-device","name":"Local and on-device use","why":"Smaller models run on laptops, phones and edge devices.","weight":{"low":0.07,"central":0.137,"high":0.25},"weightSource":{"url":"https://www.linuxfoundation.org/hubfs/LF%20Research/lfr_genai24_111924.pdf","title":"Shaping the Future of Generative AI: The Impact of Open Source Innovation","publisher":"The Linux Foundation","publishedDate":"2024-11-19","accessedDate":"2026-10-22","tier":"T4","locator":"Figure 13, p. 24: 24% deploy models locally on devices among organisations with more than 50% open-source generative-AI infrastructure; 7% among those under 25%; Figure 12: 16% of all respondents"},"shareData":"Downloads of local-format copies (GGUF, MLX) and Ollama pulls"}],"risingFloorPct":1,"validationCaseIds":[],"knownGaps":["The split of use comes from one 2024 survey that counts organisations, not volume, and covers all generative-AI models (its managed-API answers include closed models); the card uses organisations whose AI infrastructure is mostly open source as the proxy. Hosted use of open models has grown fast since (Vercel AI Gateway: open-weight models carried 71.5% of its tokens and about 23% of its spend in September 2026).","Rising growth uses one measure for every developer: monthly downloads of its open language models, including derivatives, on Hugging Face and ModelScope, from the same source in both periods.","Current commercial activity, for free models, means the developer's open models are available through at least one paid hosting provider, or a commercial deployment of them is documented.","Downloads are noisy: automated pipelines and small models inflate counts, so the ATOM Report's size-adjusted adoption measure is used as a check.","Router and gateway data cover developer traffic, not large enterprise contracts; China's hosted use is measured separately by IDC, whose totals mix open and closed models.","Use of self-hosted models inside companies is not observed directly; downloads stand in for it.","Attribution of fine-tuned copies depends on publishers tagging the base model correctly on Hugging Face.","Free use of open image, video, speech and embedding models is not measured in any sublayer."],"approval":{"approver":"Claude (delegated by Dachi)","date":"2026-10-22","sha256":"baf53a6f017635857e7234d47d52187079d61d4384a17ce50ec774dd1adb42d6","delegatedOn":"2026-10-22"}},"result":{"schema":"5lc-index-result/1","key":"models-open-weight","asOf":"2026-10-22","methodVersion":"1.1","methodSha256":"3cabed6b38514c263d8b86a494e8cc19cdbd0b2a1013543dc024192c49f7a6a6","cardSha256":"baf53a6f017635857e7234d47d52187079d61d4384a17ce50ec774dd1adb42d6","validationSha256":"409931bb458f38b02342c9bf43e61eab299e47d8679912af2c96074ec556f17e","inputsSha256":"2858435dcc525f23e6b0c22275227e6ad4c5abe7de2639973f06a01693b056e8","unscreened":[],"companies":[{"entityId":"alibaba-qwen","name":"Alibaba Qwen","reachPct":40.5902912,"importance":19.1192641068,"importanceRank":1,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":1.8735,"shareLowPct":0.7494,"shareHighPct":20,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":57.1538,"shareLowPct":22.86152,"shareHighPct":91.44608,"stale":false,"offBasis":false,"level":2,"years":0.5,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":41.5673,"shareLowPct":8.721575,"shareHighPct":74.413025,"stale":false,"offBasis":true,"level":1.33333333333,"years":0.3149802624729905,"disagreement":false}],"rising":72.2222222222,"risingEligible":true,"risingComponents":{"growth":44.4444444444,"shareChange":100},"mc":{"medianRank":1,"rankP10":1,"rankP90":1,"pTop8":1,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"deepseek","name":"DeepSeek","reachPct":16.0349713,"importance":6.04804275092,"importanceRank":2,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":47.6501,"shareLowPct":19.0601,"shareHighPct":81,"stale":false,"offBasis":false,"level":1.66666666667,"years":0.3968502629929668,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":5.4443,"shareLowPct":2.17772,"shareHighPct":8.71088,"stale":false,"offBasis":false,"level":1.33333333333,"years":0.3149802624729905,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":2.3418,"shareLowPct":1.11895,"shareHighPct":3.56465,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false}],"rising":55.5555555556,"risingEligible":true,"risingComponents":{"growth":33.3333333333,"shareChange":77.7777777778},"mc":{"medianRank":2,"rankP10":2,"rankP90":2,"pTop8":1,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"google-deepmind","name":"Google DeepMind (Gemma)","reachPct":8.614391,"importance":2.35663408875,"importanceRank":3,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.8602,"shareLowPct":0.3,"shareHighPct":4,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":11.0163,"shareLowPct":4.40652,"shareHighPct":17.62608,"stale":false,"offBasis":false,"level":1.16666666667,"years":0.2806155120779916,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":12.8326,"shareLowPct":9.33435,"shareHighPct":16.33075,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false}],"rising":50,"risingEligible":true,"risingComponents":{"growth":66.6666666667,"shareChange":33.3333333333},"mc":{"medianRank":3,"rankP10":3,"rankP90":4,"pTop8":1,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"zhipu","name":"Zhipu AI (Z.ai)","reachPct":6.3156102,"importance":1.76554265635,"importanceRank":4,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":15.5327,"shareLowPct":6.2131,"shareHighPct":24.8523,"stale":false,"offBasis":false,"level":1.16666666667,"years":0.2806155120779916,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":3.3924,"shareLowPct":1.35696,"shareHighPct":5.42784,"stale":false,"offBasis":false,"level":1.16666666667,"years":0.2806155120779916,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":1.6011,"shareLowPct":0.898125,"shareHighPct":2.304175,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false}],"rising":72.2222222222,"risingEligible":true,"risingComponents":{"growth":88.8888888889,"shareChange":55.5555555556},"mc":{"medianRank":4,"rankP10":3,"rankP90":5,"pTop8":1,"pSelected":1,"stability":"firm"},"baselineRevenueRank":1},{"entityId":"meta-superintelligence-labs","name":"Meta Superintelligence Labs (Llama)","reachPct":5.3051099,"importance":1.4167437721,"importanceRank":5,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.2393,"shareLowPct":0.0957,"shareHighPct":4,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":4.9085,"shareLowPct":1.9634,"shareHighPct":7.8536,"stale":false,"offBasis":false,"level":1.16666666667,"years":0.2806155120779916,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":16.6988,"shareLowPct":0.2848,"shareHighPct":33.1129,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false}],"rising":0,"risingEligible":true,"risingComponents":{"growth":0,"shareChange":0},"mc":{"medianRank":5,"rankP10":4,"rankP90":6,"pTop8":0.9973,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"nvidia-nemotron","name":"NVIDIA Nemotron","reachPct":3.452585,"importance":0.86314625,"importanceRank":6,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":6.5797,"shareLowPct":0.5,"shareHighPct":10.5275,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":2.7603,"shareLowPct":1.10412,"shareHighPct":4.41648,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":0.5371,"shareLowPct":0.330025,"shareHighPct":0.744275,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":88.8888888889,"risingEligible":true,"risingComponents":{"growth":88.8888888889,"shareChange":88.8888888889},"mc":{"medianRank":7,"rankP10":6,"rankP90":9,"pTop8":0.8969,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"tencent-hy","name":"Tencent Hy","reachPct":3.0192241,"importance":0.754806025,"importanceRank":7,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":10.3169,"shareLowPct":1,"shareHighPct":16.507,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":0.4062,"shareLowPct":0.16248,"shareHighPct":0.64992,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.5984,"shareLowPct":0.23936,"shareHighPct":0.95744,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":77.7777777778,"risingEligible":true,"risingComponents":{"growth":88.8888888889,"shareChange":66.6666666667},"mc":{"medianRank":8,"rankP10":6,"rankP90":9,"pTop8":0.763,"pSelected":1,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"xiaomi","name":"Xiaomi (MiMo)","reachPct":2.3936966,"importance":0.748498457455,"importanceRank":8,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":8.8488,"shareLowPct":1,"shareHighPct":27,"stale":false,"offBasis":false,"level":1.33333333333,"years":0.3149802624729905,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":0.119,"shareLowPct":0.0476,"shareHighPct":0.1904,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.0914,"shareLowPct":0.03656,"shareHighPct":0.14624,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":50,"risingEligible":true,"risingComponents":{"growth":55.5555555556,"shareChange":44.4444444444},"mc":{"medianRank":6,"rankP10":4,"rankP90":9,"pTop8":0.8458,"pSelected":0.9998,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"openai","name":"OpenAI (gpt-oss)","reachPct":1.6105565,"importance":0.402639125,"importanceRank":9,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.6239,"shareLowPct":0.2496,"shareHighPct":5,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":2.1425,"shareLowPct":0.857,"shareHighPct":5.492,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":1.1528,"shareLowPct":0.564,"shareHighPct":1.7417,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false}],"rising":16.6666666666,"risingEligible":true,"risingComponents":{"growth":22.2222222222,"shareChange":11.1111111111},"mc":{"medianRank":9,"rankP10":7,"rankP90":10,"pTop8":0.4169,"pSelected":0.9206,"stability":"firm"},"baselineRevenueRank":null},{"entityId":"mistral","name":"Mistral AI","reachPct":1.2231086,"importance":0.30577715,"importanceRank":10,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.2868,"shareLowPct":0.1147,"shareHighPct":3,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.8058,"shareLowPct":0.32232,"shareHighPct":1.28928,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":4.8406,"shareLowPct":1.54415,"shareHighPct":8.13715,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false}],"rising":16.6666666666,"risingEligible":true,"risingComponents":{"growth":11.1111111111,"shareChange":22.2222222222},"mc":{"medianRank":10,"rankP10":9,"rankP90":12,"pTop8":0.0392,"pSelected":0.5064,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"moonshot","name":"Moonshot AI","reachPct":0.9422159,"importance":0.280273723571,"importanceRank":11,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":1.8761,"shareLowPct":0.7504,"shareHighPct":5,"stale":false,"offBasis":false,"level":1.33333333333,"years":0.3149802624729905,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":0.7,"shareLowPct":0.28,"shareHighPct":1.12,"stale":false,"offBasis":false,"level":1.16666666667,"years":0.2806155120779916,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":0.2274,"shareLowPct":0.03155,"shareHighPct":0.42325,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":11,"rankP10":10,"rankP90":14,"pTop8":0.0128,"pSelected":0.269,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"hugging-face","name":"Hugging Face (SmolLM)","reachPct":0.9995489,"importance":0.249887225,"importanceRank":12,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.02,"shareLowPct":0,"shareHighPct":0.2,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":1.5318,"shareLowPct":0.61272,"shareHighPct":2.45088,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":0.5269,"shareLowPct":0.262375,"shareHighPct":0.791425,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":13,"rankP10":11,"rankP90":17,"pTop8":0.0003,"pSelected":0.0252,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"modelbest","name":"ModelBest (MiniCPM)","reachPct":0.8262402,"importance":0.228859158254,"importanceRank":13,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.05,"shareLowPct":0,"shareHighPct":0.3,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":1.2099,"shareLowPct":0.48396,"shareHighPct":1.93584,"stale":false,"offBasis":false,"level":1.16666666667,"years":0.2806155120779916,"disagreement":false},{"id":"local-device","weight":0.137,"sharePct":0.6192,"shareLowPct":0.4365,"shareHighPct":0.8019,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":14,"rankP10":12,"rankP90":17,"pTop8":0.0002,"pSelected":0.0143,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"microsoft-phi","name":"Microsoft (Phi)","reachPct":0.7207477,"importance":0.180186925,"importanceRank":14,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.0012,"shareLowPct":0.0005,"shareHighPct":0.5,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.7559,"shareLowPct":0.30236,"shareHighPct":1.20944,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":1.9371,"shareLowPct":0.047825,"shareHighPct":3.826475,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":15,"rankP10":12,"rankP90":18,"pTop8":0,"pSelected":0.0042,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"poolside","name":"Poolside (Laguna)","reachPct":0.6389497,"importance":0.159737425,"importanceRank":15,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.9214,"shareLowPct":0.05,"shareHighPct":1.4742,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":0.4044,"shareLowPct":0.16176,"shareHighPct":0.64704,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":1.1315,"shareLowPct":0.4526,"shareHighPct":1.8104,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":17,"rankP10":14,"rankP90":19,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"minimax","name":"MiniMax","reachPct":0.600197,"importance":0.15004925,"importanceRank":16,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":1.2052,"shareLowPct":0.4821,"shareHighPct":8,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":0.4312,"shareLowPct":0.17248,"shareHighPct":0.68992,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.1902,"shareLowPct":0.02715,"shareHighPct":0.35325,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":12,"rankP10":9,"rankP90":17,"pTop8":0.0276,"pSelected":0.2589,"stability":"borderline"},"baselineRevenueRank":2},{"entityId":"stepfun","name":"StepFun","reachPct":0.5014864,"importance":0.136946853638,"importanceRank":17,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":1.4486,"shareLowPct":0.5,"shareHighPct":2.5,"stale":false,"offBasis":false,"level":1.16666666667,"years":0.2806155120779916,"disagreement":false},{"id":"self-hosted","weight":0.602,"sharePct":0.1947,"shareLowPct":0.07788,"shareHighPct":0.31152,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.0452,"shareLowPct":0.01808,"shareHighPct":0.07232,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":18,"rankP10":14,"rankP90":21,"pTop8":0,"pSelected":0.0002,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"thinking-machines","name":"Thinking Machines (Inkling)","reachPct":0.5048366,"importance":0.12620915,"importanceRank":18,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.3991,"shareLowPct":0.1,"shareHighPct":2,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.5498,"shareLowPct":0.21992,"shareHighPct":0.87968,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.5087,"shareLowPct":0.20348,"shareHighPct":0.81392,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":16,"rankP10":13,"rankP90":20,"pTop8":0,"pSelected":0.0014,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"ibm-granite","name":"IBM Granite","reachPct":0.483439,"importance":0.12085975,"importanceRank":19,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.0067,"shareLowPct":0.0027,"shareHighPct":0.6,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.5604,"shareLowPct":0.22416,"shareHighPct":0.89664,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":1.0535,"shareLowPct":0.514425,"shareHighPct":1.592575,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":18,"rankP10":16,"rankP90":21,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"baidu-ernie","name":"Baidu ERNIE","reachPct":0.4159413,"importance":0.103985325,"importanceRank":20,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.2,"shareLowPct":0.02,"shareHighPct":1,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.5915,"shareLowPct":0.2366,"shareHighPct":0.9464,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.0559,"shareLowPct":0.02236,"shareHighPct":0.08944,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":19,"rankP10":16,"rankP90":22,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"liquid-ai","name":"Liquid AI","reachPct":0.2667751,"importance":0.066693775,"importanceRank":21,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.0196,"shareLowPct":0.0079,"shareHighPct":0.6,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.2152,"shareLowPct":0.08608,"shareHighPct":0.34432,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.9643,"shareLowPct":0.033425,"shareHighPct":1.895075,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":21,"rankP10":19,"rankP90":24,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"lg-ai-research","name":"LG AI Research (EXAONE)","reachPct":0.1958764,"importance":0.0489691,"importanceRank":22,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.05,"shareLowPct":0,"shareHighPct":0.3,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.2844,"shareLowPct":0.11376,"shareHighPct":0.45504,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.0848,"shareLowPct":0.0073,"shareHighPct":0.1622,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":24,"rankP10":22,"rankP90":26,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"bytedance-seed","name":"ByteDance Seed","reachPct":0.14627,"importance":0.0365675,"importanceRank":23,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.1,"shareLowPct":0.01,"shareHighPct":0.5,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.1724,"shareLowPct":0.06896,"shareHighPct":0.27584,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.1196,"shareLowPct":0.04784,"shareHighPct":0.19136,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":25,"rankP10":23,"rankP90":26,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"ant-group","name":"Ant Group (Ling and Ring)","reachPct":0.1447727,"importance":0.036193175,"importanceRank":24,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.1169,"shareLowPct":0.0468,"shareHighPct":1.5,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.0843,"shareLowPct":0.03372,"shareHighPct":0.13488,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.4636,"shareLowPct":0.18544,"shareHighPct":0.74176,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":22,"rankP10":20,"rankP90":25,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"cohere","name":"Cohere","reachPct":0.1170085,"importance":0.029252125,"importanceRank":25,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.1353,"shareLowPct":0.0541,"shareHighPct":1.5,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.0556,"shareLowPct":0.02224,"shareHighPct":0.08896,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.352,"shareLowPct":0.0822,"shareHighPct":0.6217,"stale":false,"offBasis":true,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":23,"rankP10":20,"rankP90":26,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"reflection-ai","name":"Reflection","reachPct":0.029395,"importance":0.00734875,"importanceRank":26,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.1108,"shareLowPct":0.01,"shareHighPct":0.8,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.0007,"shareLowPct":0.00028,"shareHighPct":0.00112,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.0004,"shareLowPct":0.00016,"shareHighPct":0.00064,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":27,"rankP10":25,"rankP90":27,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null},{"entityId":"meituan","name":"Meituan (LongCat)","reachPct":0.0075828,"importance":0.0018957,"importanceRank":27,"functions":[{"id":"hosted-inference","weight":0.261,"sharePct":0.0087,"shareLowPct":0.0035,"shareHighPct":1.5,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"self-hosted","weight":0.602,"sharePct":0.0078,"shareLowPct":0.00312,"shareHighPct":0.01248,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true},{"id":"local-device","weight":0.137,"sharePct":0.0045,"shareLowPct":0.0018,"shareHighPct":0.0072,"stale":false,"offBasis":false,"level":1,"years":0.25,"disagreement":false,"defaultLevel":true}],"rising":null,"risingEligible":false,"risingComponents":{},"mc":{"medianRank":26,"rankP10":22,"rankP90":27,"pTop8":0,"pSelected":0,"stability":"borderline"},"baselineRevenueRank":null}],"selection":{"importanceIds":["alibaba-qwen","deepseek","google-deepmind","zhipu","meta-superintelligence-labs","nvidia-nemotron","tencent-hy","xiaomi"],"risingIds":[],"filledFromImportance":["openai","mistral"]},"risingPools":{"growth":10,"shareChange":10,"committed":0},"weightSensitivity":[{"functionId":"hosted-inference","end":"low","entered":[],"left":[]},{"functionId":"hosted-inference","end":"high","entered":["moonshot"],"left":["mistral"]},{"functionId":"self-hosted","end":"low","entered":["moonshot"],"left":["openai"]},{"functionId":"self-hosted","end":"high","entered":["hugging-face"],"left":["mistral"]},{"functionId":"local-device","end":"low","entered":["moonshot"],"left":["mistral"]},{"functionId":"local-device","end":"high","entered":[],"left":[]}],"validation":[],"baseline":{"spearman":1,"baselineTop10NotSelected":["minimax"],"selectedNotInBaselineTop10":["alibaba-qwen","deepseek","google-deepmind","meta-superintelligence-labs","nvidia-nemotron","tencent-hy","xiaomi","openai","mistral"]},"raterAgreement":{"within":1,"weightedKappa":1,"disagreements":[]}},"inputs":{"alibaba-qwen":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"alibaba-qwen","name":"Alibaba Qwen","kind":"unit","parentId":"alibaba","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.7494,"central":1.8735,"high":20,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/qwen","title":"OpenRouter author page for qwen: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. alibaba-qwen: 2,147.5B ÷ 113,640B = 1.890% (31 models; largest: qwen/qwen3.8-flash 677.0B; qwen/qwen3.6-35b-a3b 631.0B; qwen/qwen3.8-27b 468.0B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 1.874%. Range: the 60% band around the central, widened to the screening range (2 to 20%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 5% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 31 open-weight models of this developer, 31 with a positive per-token price (highest input price $2.50 per million tokens on anthracite-org/magnum-v4-72b). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":57.1538,"central":57.1538,"high":57.1538,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. alibaba-qwen: 269.76M ÷ 472.00M = 57.154% (79,682 repositories). Largest: Qwen/Qwen3-0.6B 30.07M; Qwen/Qwen3-4B 11.70M; Qwen/Qwen3-8B 9.34M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 52.91%."},"local-device":{"low":19.1134,"central":41.5673,"high":64.0212,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. alibaba-qwen: 149.99M ÷ 234.28M = 64.021% (40,214 repositories). Largest: unsloth/Qwen3.8-27B-GGUF 6.77M; ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF 5.59M; prism-ml/Ternary-Bonsai-2-27B-gguf 4.84M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. alibaba-qwen: 185.42M ÷ 970.10M = 19.113%. Central is the mean of the two channels (64.021 and 19.113), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":17.5882,"central":17.5882,"high":17.5882,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). alibaba-qwen: 206.23B tokens ÷ 1,162.4B open-weight tokens = 17.741%, multiplied by 0.9914 as in the current estimate = 17.588% (36 models; largest: qwen/qwen3-235b-a22b-07-25 48.95B; qwen/qwen3-vl-235b-a22b-instruct 34.16B; qwen/qwen3-coder-480b-a35b-07-25 33.04B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":46.6977,"central":46.6977,"high":46.6977,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. alibaba-qwen: 97.98M ÷ 209.81M = 46.698% (25,771 repositories). Largest: Qwen/Qwen3-0.6B 7.21M; Qwen/Qwen2.5-7B-Instruct 6.04M; Qwen/Qwen3-32B 6.02M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":10.6793,"central":21.1912,"high":31.7031,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. alibaba-qwen: 10.65M ÷ 33.60M = 31.703% (19,198 repositories). Largest: unsloth/Qwen3-30B-A3B-GGUF 0.48M; lmstudio-community/Qwen3-4B-Thinking-2507-MLX-4bit 0.43M; lmstudio-community/Qwen3-4B-Thinking-2507-MLX-8bit 0.42M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. alibaba-qwen: 45.13M ÷ 422.60M = 10.679%. Central is the mean of the two channels (31.703 and 10.679), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":419754346,"central":419754346,"high":419754346,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 269.76M full-weight plus 149.99M GGUF and MLX downloads attributed to alibaba-qwen in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":108628841,"central":108628841,"high":108628841,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 97.98M full-weight plus 10.65M GGUF and MLX downloads attributed to alibaba-qwen in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":3},"self-hosted":{"levelLow":1,"levelHigh":3},"local-device":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"6ac4e80235d6cce93115b1944ed567319dfd22be8e3993955dbced2021157641"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"6ac4e80235d6cce93115b1944ed567319dfd22be8e3993955dbced2021157641"}}},"deepseek":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"deepseek","name":"DeepSeek","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":19.0601,"central":47.6501,"high":81,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/deepseek","title":"OpenRouter author page for deepseek: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. deepseek: 54,619.3B ÷ 113,640B = 48.063% (14 models; largest: deepseek/deepseek-v4.1-flash 23,600.0B; deepseek/deepseek-v4.1-flash-1017 21,900.0B; deepseek/deepseek-v4-flash-0731 4,210.0B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 47.650%. Range: the 60% band around the central, widened to the screening range (31 to 81%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 52% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 15 open-weight models of this developer, 15 with a positive per-token price (highest input price $0.96 per million tokens on deepseek/deepseek-v4-pro). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":5.4443,"central":5.4443,"high":5.4443,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. deepseek: 25.70M ÷ 472.00M = 5.444% (4,455 repositories). Largest: deepseek-ai/DeepSeek-V3.2 5.15M; deepseek-ai/DeepSeek-V4-Flash-0731 4.33M; deepseek-ai/DeepSeek-OCR 2.08M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 5.04%."},"local-device":{"low":1.7044,"central":2.3418,"high":2.9792,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. deepseek: 6.98M ÷ 234.28M = 2.979% (3,524 repositories). Largest: antirez/deepseek-v4.1-flash-gguf 1.63M; antirez/deepseek-v4-gguf 1.39M; huihui-ai/Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF 0.35M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. deepseek: 16.53M ÷ 970.10M = 1.704%. Central is the mean of the two channels (2.979 and 1.704), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":36.7481,"central":36.7481,"high":36.7481,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). deepseek: 430.88B tokens ÷ 1,162.4B open-weight tokens = 37.067%, multiplied by 0.9914 as in the current estimate = 36.748% (13 models; largest: deepseek/deepseek-chat-v3-0324 125.53B; tngtech/deepseek-r1t2-chimera 83.79B; deepseek/deepseek-chat-v3.1 78.83B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":4.8895,"central":4.8895,"high":4.8895,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. deepseek: 10.26M ÷ 209.81M = 4.889% (2,787 repositories). Largest: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B 2.61M; deepseek-ai/DeepSeek-R1-Distill-Llama-8B 0.82M; deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B 0.69M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":1.3411,"central":3.5688,"high":5.7965,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. deepseek: 1.95M ÷ 33.60M = 5.796% (3,094 repositories). Largest: lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit 0.24M; lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit 0.23M; mradermacher/DeepSeek-V2-Lite-GGUF 0.14M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. deepseek: 5.67M ÷ 422.60M = 1.341%. Central is the mean of the two channels (5.796 and 1.341), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":32676801,"central":32676801,"high":32676801,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 25.70M full-weight plus 6.98M GGUF and MLX downloads attributed to deepseek in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":12206353,"central":12206353,"high":12206353,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 10.26M full-weight plus 1.95M GGUF and MLX downloads attributed to deepseek in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":3},"self-hosted":{"levelLow":1,"levelHigh":3},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"21c26c61546649fd811d20cc4fa812d50a5f68530f5878b45ae30dc6e84f82cd"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"21c26c61546649fd811d20cc4fa812d50a5f68530f5878b45ae30dc6e84f82cd"}}},"google-deepmind":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"google-deepmind","name":"Google DeepMind (Gemma)","kind":"unit","parentId":"alphabet","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.3,"central":0.8602,"high":4,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/google","title":"OpenRouter author page for google: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. google-deepmind: 986.0B ÷ 113,640B = 0.868% (8 models; largest: google/gemma-4-31b-it 567.0B; google/gemma-4-26b-a4b-it 372.0B; google/gemma-3-27b-it 31.7B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.860%. Range: the 60% band around the central, widened to the screening range (0.3 to 4%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 1.2% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 8 open-weight models of this developer, 6 with a positive per-token price (highest input price $0.65 per million tokens on google/gemma-2-27b-it). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":11.0163,"central":11.0163,"high":11.0163,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. google-deepmind: 52.00M ÷ 472.00M = 11.016% (24,701 repositories). Largest: google/gemma-4-26B-A4B-it 10.74M; google/gemma-4-31B-it 9.11M; google/gemma-4-E4B-it 4.21M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 10.68%."},"local-device":{"low":12.5425,"central":12.8326,"high":13.1226,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. google-deepmind: 30.74M ÷ 234.28M = 13.123% (13,246 repositories). Largest: unsloth/gemma-4-12b-it-GGUF 1.64M; lmstudio-community/gemma-4-E4B-it-MLX-4bit 1.14M; lmstudio-community/gemma-4-E4B-it-MLX-8bit 1.11M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. google-deepmind: 121.67M ÷ 970.10M = 12.543%. Central is the mean of the two channels (13.123 and 12.543), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":3.5416,"central":3.5416,"high":3.5416,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). google-deepmind: 41.53B tokens ÷ 1,162.4B open-weight tokens = 3.572%, multiplied by 0.9914 as in the current estimate = 3.542% (7 models; largest: google/gemma-3-27b-it 25.34B; google/gemma-3-12b-it 6.83B; google/gemma-3n-e4b-it 6.33B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":8.1069,"central":8.1069,"high":8.1069,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. google-deepmind: 17.01M ÷ 209.81M = 8.107% (11,736 repositories). Largest: google/gemma-3-1b-it 7.28M; google/gemma-3-27b-it 1.17M; google/paligemma2-3b-pt-224 1.16M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":8.5117,"central":9.7303,"high":10.949,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. google-deepmind: 3.68M ÷ 33.60M = 10.949% (6,869 repositories). Largest: google/gemma-3-27b-it-qat-q4_0-gguf 0.19M; lmstudio-community/gemma-3n-E4B-it-MLX-4bit 0.19M; google/gemma-2b 0.18M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. google-deepmind: 35.97M ÷ 422.60M = 8.512%. Central is the mean of the two channels (10.949 and 8.512), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":82740502,"central":82740502,"high":82740502,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 52.00M full-weight plus 30.74M GGUF and MLX downloads attributed to google-deepmind in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":20688292,"central":20688292,"high":20688292,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 17.01M full-weight plus 3.68M GGUF and MLX downloads attributed to google-deepmind in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":3},"self-hosted":{"levelLow":1,"levelHigh":3},"local-device":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"651d1b253185374ca11a71b659111140fbf441f22f0bef445766294495ef358f"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"651d1b253185374ca11a71b659111140fbf441f22f0bef445766294495ef358f"}}},"zhipu":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"zhipu","name":"Zhipu AI (Z.ai)","kind":"public","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":6.2131,"central":15.5327,"high":24.8523,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/z-ai","title":"OpenRouter author page for z-ai: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. zhipu: 17,804.4B ÷ 113,640B = 15.667% (14 models; largest: z-ai/glm-5.3-flash 11,500.0B; z-ai/glm-5.3 5,190.0B; z-ai/glm-5.2 970.0B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 15.533%. Range: the 60% band around the central, widened to the screening range (8 to 18%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 12% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 14 open-weight models of this developer, 14 with a positive per-token price (highest input price $1.40 per million tokens on z-ai/glm-5.1). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":3.3924,"central":3.3924,"high":3.3924,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. zhipu: 16.01M ÷ 472.00M = 3.392% (1,055 repositories). Largest: zai-org/GLM-5.3-Flash 6.57M; zai-org/GLM-5.3 1.86M; zai-org/GLM-4.7-Flash 1.22M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 3.14%."},"local-device":{"low":1.2984,"central":1.6011,"high":1.9039,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. zhipu: 4.46M ÷ 234.28M = 1.904% (1,256 repositories). Largest: unsloth/GLM-5.3-Flash-GGUF 0.96M; BoldingBuilds/orcarouter_GLM-5.3-Flash-Uncensored-GGUF 0.48M; unsloth/GLM-5.3-GGUF 0.42M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. zhipu: 12.60M ÷ 970.10M = 1.298%. Central is the mean of the two channels (1.904 and 1.298), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":11.3081,"central":11.3081,"high":11.3081,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). zhipu: 132.59B tokens ÷ 1,162.4B open-weight tokens = 11.406%, multiplied by 0.9914 as in the current estimate = 11.308% (7 models; largest: z-ai/glm-4.6 77.12B; z-ai/glm-4.5-air 46.50B; z-ai/glm-4.5 8.27B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":0.8388,"central":0.8388,"high":0.8388,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. zhipu: 1.76M ÷ 209.81M = 0.839% (203 repositories). Largest: zai-org/GLM-4.5V-FP8 0.42M; zai-org/GLM-4.5-Air 0.36M; zai-org/GLM-4.1V-9B-Thinking 0.28M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0.0378,"central":0.5663,"high":1.0948,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. zhipu: 0.37M ÷ 33.60M = 1.095% (486 repositories). Largest: unsloth/GLM-4.6-GGUF 0.08M; unsloth/GLM-4.5-Air-GGUF 0.05M; bartowski/zai-org_GLM-4.6-GGUF 0.02M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. zhipu: 0.16M ÷ 422.60M = 0.038%. Central is the mean of the two channels (1.095 and 0.038), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":20472337,"central":20472337,"high":20472337,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 16.01M full-weight plus 4.46M GGUF and MLX downloads attributed to zhipu in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":2127739,"central":2127739,"high":2127739,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 1.76M full-weight plus 0.37M GGUF and MLX downloads attributed to zhipu in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"baselineRevenueUsd":{"low":104860000,"central":104860000,"high":104860000,"unit":"USD","basis":"Total revenue, calendar year, US dollars","period":"CY2025","periodEnd":"2025-12-31","source":{"url":"https://whbl.com/2026/03/31/chinas-zhipu-posts-132-rise-in-annual-revenue-on-ai-boom/","title":"China's Zhipu posts 132% rise in annual revenue on AI boom","publisher":"Reuters (syndicated by WHBL)","publishedDate":"2026-03-31","accessedDate":"2026-10-22","tier":"T3","locator":"segment figures: on-premise deployment 533.9 million yuan, cloud API 190.4 million yuan; $1 = 6.9074 yuan"},"note":"Reuters reports Zhipu's 2025 revenue by segment only: 533.9 million yuan on-premise deployment plus 190.4 million yuan cloud API services = 724.3 million yuan, which at the article's rate of 6.9074 yuan per US dollar is $104.86 million. The article does not state a total, so revenue outside these two segments, if any, is not shown. Zhipu is the entity scored, so this is its own revenue, mostly from GLM models it also publishes as open weights."},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":3},"self-hosted":{"levelLow":1,"levelHigh":3},"local-device":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"60c8067f3435b02ba58a0963e05114ad6a33d1b2623873a72473112bec698d52"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"60c8067f3435b02ba58a0963e05114ad6a33d1b2623873a72473112bec698d52"}}},"meta-superintelligence-labs":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"meta-superintelligence-labs","name":"Meta Superintelligence Labs (Llama)","kind":"unit","parentId":"meta","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.0957,"central":0.2393,"high":4,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/meta-llama","title":"OpenRouter author page for meta-llama: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. meta-superintelligence-labs: 274.3B ÷ 113,640B = 0.241% (16 models; largest: meta/muse-spark-open 117.0B; meta-llama/llama-3.1-8b-instruct 60.1B; meta-llama/llama-3.3-70b-instruct 35.7B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.239%. Range: the 60% band around the central, widened to the screening range (0.1 to 4%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.8% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 16 open-weight models of this developer, 16 with a positive per-token price (highest input price $1.25 per million tokens on meta/muse-spark-open). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":4.9085,"central":4.9085,"high":4.9085,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. meta-superintelligence-labs: 23.17M ÷ 472.00M = 4.909% (36,701 repositories). Largest: meta-llama/Llama-3.1-8B-Instruct 5.65M; meta-llama/Llama-3.2-1B-Instruct 5.49M; meta-llama/Llama-3.2-3B-Instruct 1.22M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 5.99%. The range is the T5 tier width around this one figure. Another observation, on a different basis and not used as a bound: the ATOM Report (Lambert and Brand, https://arxiv.org/abs/2604.07190, v2, T4) gives Llama 476.0M cumulative downloads by March 2026 (section 5.1) out of 2.04 billion cumulative tracked downloads (page 6), about 23.3% by division; the report prints the two totals, not the share. These are all-time totals for the report's tracked models, not the latest month, and the report states no January 2026 download share for Llama."},"local-device":{"low":4.4595,"central":16.6988,"high":28.9382,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. meta-superintelligence-labs: 10.45M ÷ 234.28M = 4.459% (22,757 repositories). Largest: meta-models/Muse-Glimmer-30B-GGUF 0.79M; bartowski/Meta-Llama-3.1-8B-Instruct-GGUF 0.50M; unsloth/Muse-Glimmer-30B-GGUF 0.44M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. meta-superintelligence-labs: 280.73M ÷ 970.10M = 28.938%. Central is the mean of the two channels (4.459 and 28.938), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":7.8069,"central":7.8069,"high":7.8069,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). meta-superintelligence-labs: 91.54B tokens ÷ 1,162.4B open-weight tokens = 7.875%, multiplied by 0.9914 as in the current estimate = 7.807% (37 models; largest: meta-llama/llama-3.3-70b-instruct 23.76B; meta-llama/llama-4-maverick-17b-128e-instruct 21.36B; meta-llama/llama-3.1-8b-instruct 12.62B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":15.2645,"central":15.2645,"high":15.2645,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. meta-superintelligence-labs: 32.03M ÷ 209.81M = 15.264% (25,881 repositories). Largest: meta-llama/Llama-3.1-8B-Instruct 5.85M; meta-llama/Llama-3.2-1B-Instruct 4.95M; meta-llama/Llama-3.2-3B-Instruct 1.89M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":13.732,"central":27.7959,"high":41.8599,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. meta-superintelligence-labs: 4.61M ÷ 33.60M = 13.732% (20,941 repositories). Largest: bartowski/Llama-3.2-3B-Instruct-GGUF 0.25M; MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF 0.12M; MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF 0.10M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. meta-superintelligence-labs: 176.90M ÷ 422.60M = 41.860%. Central is the mean of the two channels (13.732 and 41.860), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":33615810,"central":33615810,"high":33615810,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 23.17M full-weight plus 10.45M GGUF and MLX downloads attributed to meta-superintelligence-labs in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":36640549,"central":36640549,"high":36640549,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 32.03M full-weight plus 4.61M GGUF and MLX downloads attributed to meta-superintelligence-labs in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":3},"self-hosted":{"levelLow":1,"levelHigh":3},"local-device":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c3096db93c2b6d0044684329ce3a0c3dd6a061750f00524eda8044f54b89c9fc"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c3096db93c2b6d0044684329ce3a0c3dd6a061750f00524eda8044f54b89c9fc"}}},"nvidia-nemotron":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"nvidia-nemotron","name":"NVIDIA Nemotron","kind":"unit","parentId":"nvidia","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.5,"central":6.5797,"high":10.5275,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/nvidia","title":"OpenRouter author page for nvidia: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. nvidia-nemotron: 7,542.0B ÷ 113,640B = 6.637% (12 models; largest: nvidia/nemotron-3-ultra-550b-a55b:free 4,680.0B; nvidia/nemotron-3.5-super-vl:free 1,390.0B; nvidia/nemotron-3.5-lightning:free 816.0B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 6.580%. Range: the 60% band around the central, widened to the screening range (0.5 to 4%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 2.5% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 12 open-weight models of this developer, 6 with a positive per-token price (highest input price $0.60 per million tokens on nvidia/nemotron-3-ultra-550b-a55b). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":2.7603,"central":2.7603,"high":2.7603,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. nvidia-nemotron: 13.03M ÷ 472.00M = 2.760% (669 repositories). Largest: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 1.70M; nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 1.57M; nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 1.16M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 2.56%."},"local-device":{"low":0.4643,"central":0.5371,"high":0.61,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. nvidia-nemotron: 1.09M ÷ 234.28M = 0.464% (949 repositories). Largest: ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF 0.18M; nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF 0.05M; MaziyarPanahi/NVIDIA-Nemotron-Nano-12B-v2-GGUF 0.05M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. nvidia-nemotron: 5.92M ÷ 970.10M = 0.610%. Central is the mean of the two channels (0.464 and 0.610), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":0.0728,"central":0.0728,"high":0.0728,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). nvidia-nemotron: 0.85B tokens ÷ 1,162.4B open-weight tokens = 0.073%, multiplied by 0.9914 as in the current estimate = 0.073% (1 models; largest: nvidia/nemotron-nano-9b-v2 0.85B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":0.235,"central":0.235,"high":0.235,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. nvidia-nemotron: 0.49M ÷ 209.81M = 0.235% (44 repositories). Largest: nvidia/NVIDIA-Nemotron-Nano-9B-v2 0.30M; nvidia/NVIDIA-Nemotron-Nano-12B-v2 0.10M; nvidia/NVIDIA-Nemotron-Nano-9B-v2-Base 0.02M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0.0255,"central":0.126,"high":0.2265,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. nvidia-nemotron: 0.08M ÷ 33.60M = 0.226% (229 repositories). Largest: bartowski/nvidia_NVIDIA-Nemotron-Nano-9B-v2-GGUF 0.01M; bartowski/TheDrummer_Snowpiercer-15B-v3-GGUF 0.01M; bartowski/nvidia_NVIDIA-Nemotron-Nano-12B-v2-GGUF 0.01M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. nvidia-nemotron: 0.11M ÷ 422.60M = 0.026%. Central is the mean of the two channels (0.226 and 0.026), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":14116144,"central":14116144,"high":14116144,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 13.03M full-weight plus 1.09M GGUF and MLX downloads attributed to nvidia-nemotron in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":569134,"central":569134,"high":569134,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 0.49M full-weight plus 0.08M GGUF and MLX downloads attributed to nvidia-nemotron in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":3},"self-hosted":{"levelLow":1,"levelHigh":3},"local-device":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"81d2d820018ad75821722c83e0e1d13440d6b4fdd739b91b8ded7b9f7ea81610"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"81d2d820018ad75821722c83e0e1d13440d6b4fdd739b91b8ded7b9f7ea81610"}}},"tencent-hy":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"tencent-hy","name":"Tencent Hy","kind":"unit","parentId":"tencent","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":1,"central":10.3169,"high":16.507,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/tencent","title":"OpenRouter author page for tencent: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. tencent-hy: 11,825.8B ÷ 113,640B = 10.406% (7 models; largest: tencent/hy4-preview 8,750.0B; tencent/hy3 3,070.0B; tencent/hy3-preview 4.5B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 10.317%. Range: the 60% band around the central, widened to the screening range (1 to 10%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 5% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 7 open-weight models of this developer, 7 with a positive per-token price (highest input price $0.83 per million tokens on tencent/hy4-preview). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.4062,"central":0.4062,"high":0.4062,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. tencent-hy: 1.92M ÷ 472.00M = 0.406% (129 repositories). Largest: tencent/HunyuanOCR 0.89M; tencent/Hy3 0.72M; tencent/Hy3-preview 0.10M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.38%."},"local-device":{"low":0.5984,"central":0.5984,"high":0.5984,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. tencent-hy: 1.40M ÷ 234.28M = 0.598% (265 repositories). Largest: tencent/Hy-MT2-7B-GGUF 0.79M; tencent/Hy-MT2-1.8B-GGUF 0.30M; AngelSlim/Hy3-GGUF 0.22M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"priorShares":{"hosted-inference":{"low":0.0018,"central":0.0018,"high":0.0018,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). tencent-hy: 0.02B tokens ÷ 1,162.4B open-weight tokens = 0.002%, multiplied by 0.9914 as in the current estimate = 0.002% (1 models; largest: tencent/hunyuan-a13b-instruct 0.02B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":0.0179,"central":0.0179,"high":0.0179,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. tencent-hy: 0.04M ÷ 209.81M = 0.018% (42 repositories). Largest: tencent/Hunyuan-A13B-Instruct 0.01M; tencent/Hunyuan-7B-Instruct 0.01M; tencent/Hunyuan-A13B-Instruct-FP8 0.01M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0.0967,"central":0.0967,"high":0.0967,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. tencent-hy: 0.03M ÷ 33.60M = 0.097% (91 repositories). Largest: bartowski/tencent_Hunyuan-1.8B-Instruct-GGUF 0.01M; bartowski/tencent_Hunyuan-7B-Instruct-GGUF 0.01M; tencent/Hunyuan-A13B-Instruct-GGUF 0.00M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"activityGrowth":{"current":{"low":3318867,"central":3318867,"high":3318867,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 1.92M full-weight plus 1.40M GGUF and MLX downloads attributed to tencent-hy in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":69947,"central":69947,"high":69947,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 0.04M full-weight plus 0.03M GGUF and MLX downloads attributed to tencent-hy in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"cc38446913c421aba6b0aec254c9cb0e4e914a100395c3670fb61ef31518b9f0"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"cc38446913c421aba6b0aec254c9cb0e4e914a100395c3670fb61ef31518b9f0"}}},"xiaomi":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"xiaomi","name":"Xiaomi (MiMo)","kind":"public","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":1,"central":8.8488,"high":27,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/xiaomi","title":"OpenRouter author page for xiaomi: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. xiaomi: 10,143.0B ÷ 113,640B = 8.926% (4 models; largest: xiaomi/mimo-v2.6-flash 8,040.0B; xiaomi/mimo-v2.6-pro 1,600.0B; xiaomi/mimo-v2.5 374.0B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 8.849%. Range: the 60% band around the central, widened to the screening range (1 to 27%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 7% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 4 open-weight models of this developer, 4 with a positive per-token price (highest input price $0.43 per million tokens on xiaomi/mimo-v2.6-pro). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.119,"central":0.119,"high":0.119,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. xiaomi: 0.56M ÷ 472.00M = 0.119% (125 repositories). Largest: XiaomiMiMo/MiMo-V2.5 0.17M; XiaomiMiMo/MiMo-V2.6-Pro-RL 0.10M; XiaomiMiMo/MiMo-V2.6-Flash-RL 0.10M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.11%."},"local-device":{"low":0.0914,"central":0.0914,"high":0.0914,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. xiaomi: 0.21M ÷ 234.28M = 0.091% (172 repositories). Largest: QuantaPlanta/MiMo-V2.6-Flash-MOPD-MXFP4-GGUF 0.05M; unsloth/MiMo-V2-Flash-GGUF 0.02M; ggml-org/MiMo-V2.6-Flash-RL-GGUF 0.02M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"priorShares":{"hosted-inference":{"low":0,"central":0,"high":0,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). No model of xiaomi appears among the 284 models in the week's data, so the observed share is 0 (a complete listing, not a missing figure). The ATOM Report (https://arxiv.org/abs/2604.07190, T4) says Xiaomi's MiMo-V2-Flash, which drove its OpenRouter usage, was released in December 2025 and rose from zero to 27.2% of OpenRouter's open-model token share by January 2026. Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":0.0887,"central":0.0887,"high":0.0887,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. xiaomi: 0.19M ÷ 209.81M = 0.089% (31 repositories). Largest: XiaomiMiMo/MiMo-7B-Base 0.06M; XiaomiMiMo/MiMo-VL-7B-SFT 0.04M; XiaomiMiMo/MiMo-7B-RL 0.04M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0.0306,"central":0.0306,"high":0.0306,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. xiaomi: 0.01M ÷ 33.60M = 0.031% (46 repositories). Largest: bartowski/XiaomiMiMo_MiMo-VL-7B-RL-2508-GGUF 0.00M; bartowski/XiaomiMiMo_MiMo-VL-7B-SFT-2508-GGUF 0.00M; bartowski/Aurore-Reveil_Koto-Small-7B-IT-GGUF 0.00M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"activityGrowth":{"current":{"low":775840,"central":775840,"high":775840,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 0.56M full-weight plus 0.21M GGUF and MLX downloads attributed to xiaomi in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":196473,"central":196473,"high":196473,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 0.19M full-weight plus 0.01M GGUF and MLX downloads attributed to xiaomi in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"b344c47b081d1ee292291033c433997f4dcd1bafc70e083477c55aaca5171621"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"b344c47b081d1ee292291033c433997f4dcd1bafc70e083477c55aaca5171621"}}},"openai":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"openai","name":"OpenAI (gpt-oss)","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.2496,"central":0.6239,"high":5,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/openai","title":"OpenRouter author page for openai: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. openai: 715.2B ÷ 113,640B = 0.629% (4 models; largest: openai/gpt-oss-120b 512.0B; openai/gpt-oss-20b 202.0B; openai/gpt-oss-20b:batch 0.8B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.624%. Range: the 60% band around the central, widened to the screening range (0.5 to 5%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 1.5% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 4 open-weight models of this developer, 4 with a positive per-token price (highest input price $0.15 per million tokens on openai/gpt-oss-120b). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.857,"central":2.1425,"high":5.492,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. openai: 10.11M ÷ 472.00M = 2.142% (3,796 repositories). Largest: openai/gpt-oss-20b 5.98M; openai/gpt-oss-120b 3.31M; FastFlowLM/GPT-OSS-20B-NPU2 0.39M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 5.49%. The high end includes the pre-2023-07-18 GPT-2 repositories (openai-community), which the rule leaves out."},"local-device":{"low":0.8522,"central":1.1528,"high":1.4535,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. openai: 2.00M ÷ 234.28M = 0.852% (1,111 repositories). Largest: unsloth/gpt-oss-20b-GGUF 0.53M; mlx-community/gpt-oss-20b-MXFP4-Q8 0.20M; ggml-org/gpt-oss-20b-GGUF 0.20M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. openai: 14.10M ÷ 970.10M = 1.453%. Central is the mean of the two channels (0.852 and 1.453), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":10.4228,"central":10.4228,"high":10.4228,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). openai: 122.21B tokens ÷ 1,162.4B open-weight tokens = 10.513%, multiplied by 0.9914 as in the current estimate = 10.423% (2 models; largest: openai/gpt-oss-120b 79.16B; openai/gpt-oss-20b 43.05B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":4.5829,"central":4.5829,"high":4.5829,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. openai: 9.62M ÷ 209.81M = 4.583% (2,370 repositories). Largest: openai/gpt-oss-20b 5.58M; openai/gpt-oss-120b 3.25M; unsloth/gpt-oss-20b-BF16 0.39M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0.8045,"central":3.373,"high":5.9415,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. openai: 2.00M ÷ 33.60M = 5.941% (562 repositories). Largest: mlx-community/gpt-oss-20b-MXFP4-Q8 0.81M; unsloth/gpt-oss-20b-GGUF 0.21M; lmstudio-community/gpt-oss-20b-GGUF 0.20M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. openai: 3.40M ÷ 422.60M = 0.805%. Central is the mean of the two channels (5.941 and 0.805), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":12108965,"central":12108965,"high":12108965,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 10.11M full-weight plus 2.00M GGUF and MLX downloads attributed to openai in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":11611880,"central":11611880,"high":11611880,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 9.62M full-weight plus 2.00M GGUF and MLX downloads attributed to openai in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":3},"self-hosted":{"levelLow":1,"levelHigh":3},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"3e6e6300e45e22124858b16e02fd94059bbe7220cac07968ab8a4e60b85f8991"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"3e6e6300e45e22124858b16e02fd94059bbe7220cac07968ab8a4e60b85f8991"}}},"mistral":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"mistral","name":"Mistral AI","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.1147,"central":0.2868,"high":3,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/mistralai","title":"OpenRouter author page for mistralai: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. mistral: 328.7B ÷ 113,640B = 0.289% (16 models; largest: mistralai/mistral-nemo 267.0B; mistralai/mistral-small-3.2-24b-instruct 28.9B; mistralai/mistral-small-24b-instruct-2501 11.5B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.287%. Range: the 60% band around the central, widened to the screening range (0.3 to 3%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 1.0% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 17 open-weight models of this developer, 17 with a positive per-token price (highest input price $2.00 per million tokens on mistralai/mixtral-8x22b-instruct). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.8058,"central":0.8058,"high":0.8058,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. mistral: 3.80M ÷ 472.00M = 0.806% (13,109 repositories). Largest: mistralai/Mistral-7B-Instruct-v0.2 1.66M; unsloth/mistral-7b-v0.3-bnb-4bit 0.37M; mistralai/Mistral-7B-v0.1 0.33M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.75%."},"local-device":{"low":2.7543,"central":4.8406,"high":6.927,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. mistral: 6.45M ÷ 234.28M = 2.754% (10,534 repositories). Largest: MaziyarPanahi/Mistral-7B-v0.1-GGUF 0.15M; MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF 0.12M; MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF 0.11M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. mistral: 67.20M ÷ 970.10M = 6.927%. Central is the mean of the two channels (2.754 and 6.927), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":7.7923,"central":7.7923,"high":7.7923,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). mistral: 91.37B tokens ÷ 1,162.4B open-weight tokens = 7.860%, multiplied by 0.9914 as in the current estimate = 7.792% (30 models; largest: mistralai/mistral-nemo 42.40B; mistralai/mistral-small-3.2-24b-instruct-2506 13.94B; mistralai/mistral-small-24b-instruct-2501 11.25B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":2.125,"central":2.125,"high":2.125,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. mistral: 4.46M ÷ 209.81M = 2.125% (10,294 repositories). Largest: mistralai/Mistral-7B-Instruct-v0.2 1.80M; unsloth/mistral-7b-v0.3-bnb-4bit 0.40M; mistralai/Mistral-7B-v0.1 0.36M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":7.3748,"central":12.044,"high":16.7133,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. mistral: 5.62M ÷ 33.60M = 16.713% (8,887 repositories). Largest: lmstudio-community/Magistral-Small-2509-MLX-4bit 0.65M; lmstudio-community/Magistral-Small-2509-MLX-8bit 0.64M; lmstudio-community/Magistral-Small-2509-MLX-6bit 0.64M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. mistral: 31.17M ÷ 422.60M = 7.375%. Central is the mean of the two channels (16.713 and 7.375), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":10256129,"central":10256129,"high":10256129,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 3.80M full-weight plus 6.45M GGUF and MLX downloads attributed to mistral in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":10074813,"central":10074813,"high":10074813,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 4.46M full-weight plus 5.62M GGUF and MLX downloads attributed to mistral in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":3},"self-hosted":{"levelLow":1,"levelHigh":3},"local-device":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"ddf5b40638088c0fed8d55203e7b68e86a6507ea4f000ef70a68f2ff01a430a9"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"ddf5b40638088c0fed8d55203e7b68e86a6507ea4f000ef70a68f2ff01a430a9"}}},"moonshot":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"moonshot","name":"Moonshot AI","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.7504,"central":1.8761,"high":5,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/moonshotai","title":"OpenRouter author page for moonshotai: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. moonshot: 2,150.5B ÷ 113,640B = 1.892% (8 models; largest: moonshotai/kimi-k3 1,920.0B; moonshotai/kimi-k2.6 135.0B; moonshotai/kimi-k2.7-code 55.1B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 1.876%. Range: the 60% band around the central, widened to the screening range (1.5 to 5%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 3.2% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 8 open-weight models of this developer, 8 with a positive per-token price (highest input price $2.28 per million tokens on moonshotai/kimi-k3:batch). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.7,"central":0.7,"high":0.7,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. moonshot: 3.30M ÷ 472.00M = 0.700% (151 repositories). Largest: moonshotai/Kimi-K3 1.05M; moonshotai/Kimi-K2.6 0.38M; moonshotai/Kimi-K2-Instruct 0.35M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.65%."},"local-device":{"low":0.0884,"central":0.2274,"high":0.3664,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. moonshot: 0.86M ÷ 234.28M = 0.366% (212 repositories). Largest: unsloth/Kimi-K3-GGUF 0.31M; unsloth/Kimi-K2.7-Code-GGUF 0.26M; mlx-community/Kimi-K2.5 0.12M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. moonshot: 0.86M ÷ 970.10M = 0.088%. Central is the mean of the two channels (0.366 and 0.088), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":2.6331,"central":2.6331,"high":2.6331,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). moonshot: 30.87B tokens ÷ 1,162.4B open-weight tokens = 2.656%, multiplied by 0.9914 as in the current estimate = 2.633% (3 models; largest: moonshotai/kimi-k2-0905 22.33B; moonshotai/kimi-k2 8.04B; moonshotai/kimi-dev-72b 0.50B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":0.2129,"central":0.2129,"high":0.2129,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. moonshot: 0.45M ÷ 209.81M = 0.213% (25 repositories). Largest: moonshotai/Kimi-VL-A3B-Instruct 0.14M; moonshotai/Kimi-K2-Instruct 0.11M; moonshotai/Kimi-VL-A3B-Thinking-2506 0.09M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0.0013,"central":0.0739,"high":0.1465,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. moonshot: 0.05M ÷ 33.60M = 0.147% (50 repositories). Largest: unsloth/Kimi-K2-Instruct-GGUF 0.03M; unsloth/Kimi-K2-Instruct-0905-GGUF 0.01M; ggml-org/Kimi-VL-A3B-Thinking-2506-GGUF 0.00M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. moonshot: 0.01M ÷ 422.60M = 0.001%. Central is the mean of the two channels (0.147 and 0.001), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":4162068,"central":4162068,"high":4162068,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 3.30M full-weight plus 0.86M GGUF and MLX downloads attributed to moonshot in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":495922,"central":495922,"high":495922,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 0.45M full-weight plus 0.05M GGUF and MLX downloads attributed to moonshot in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":3},"self-hosted":{"levelLow":1,"levelHigh":3},"local-device":{"levelLow":1,"levelHigh":3}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"0e20685cd31f25a4c33851e85d939e284ddb623e389568ddaa479d6165eefea9"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"0e20685cd31f25a4c33851e85d939e284ddb623e389568ddaa479d6165eefea9"}}},"hugging-face":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"hugging-face","name":"Hugging Face (SmolLM)","kind":"private","commercialActivityVerified":false,"shares":{"hosted-inference":{"low":0,"central":0.02,"high":0.2,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter listing at 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/api/v1/models","title":"OpenRouter model list (public API; models with a Hugging Face weights id and their per-token prices)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Entries with a hugging_face_id and a positive pricing.prompt; live listing, no publication date shown"},"note":"5LayerCake estimate (T5). Method: shares of tokens served through gateways, re-based to open-weight models only (a model counts as open when OpenRouter's model list gives it a Hugging Face weights id). Allowance only. No SmolLM model is listed on OpenRouter and no paid hosting was found. Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":1.5318,"central":1.5318,"high":1.5318,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. hugging-face: 7.23M ÷ 472.00M = 1.532% (3,166 repositories). Largest: HuggingFaceTB/SmolLM2-135M 1.68M; HuggingFaceTB/SmolVLM2-500M-Video-Instruct 1.39M; HuggingFaceTB/SmolLM2-135M-Instruct 1.37M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 1.42%."},"local-device":{"low":0.3941,"central":0.5269,"high":0.6597,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. hugging-face: 0.92M ÷ 234.28M = 0.394% (1,240 repositories). Largest: webAI-Official/TwIL-LM3 0.25M; unsloth/SmolLM2-135M-Instruct-GGUF 0.07M; bartowski/SmolLM2-135M-Instruct-GGUF 0.05M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. hugging-face: 6.40M ÷ 970.10M = 0.660%. Central is the mean of the two channels (0.394 and 0.660), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":0,"central":0,"high":0,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). No model of hugging-face appears among the 284 models in the week's data, so the observed share is 0 (a complete listing, not a missing figure). SmolLM models are not listed on OpenRouter. Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range. No model of this developer is listed on OpenRouter now, so the current figure is an allowance, not a measurement."},"self-hosted":{"low":1.4534,"central":1.4534,"high":1.4534,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. hugging-face: 3.05M ÷ 209.81M = 1.453% (1,923 repositories). Largest: HuggingFaceTB/SmolLM2-135M 0.62M; HuggingFaceTB/SmolVLM2-256M-Video-Instruct 0.46M; HuggingFaceTB/SmolLM-135M 0.42M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0.5612,"central":0.7257,"high":0.8903,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. hugging-face: 0.30M ÷ 33.60M = 0.890% (873 repositories). Largest: ggml-org/SmolVLM-500M-Instruct-GGUF 0.12M; prithivMLmods/SmolLM2-135M-GGUF 0.02M; Felladrin/gguf-smollm-360M-instruct-add-basics 0.01M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. hugging-face: 2.37M ÷ 422.60M = 0.561%. Central is the mean of the two channels (0.890 and 0.561), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":8153192,"central":8153192,"high":8153192,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 7.23M full-weight plus 0.92M GGUF and MLX downloads attributed to hugging-face in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":3348592,"central":3348592,"high":3348592,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 3.05M full-weight plus 0.30M GGUF and MLX downloads attributed to hugging-face in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"4b05ccab164178ab73b7c686c5efb008ff4916be605160b55830c43146c79f29"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"4b05ccab164178ab73b7c686c5efb008ff4916be605160b55830c43146c79f29"}}},"modelbest":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"modelbest","name":"ModelBest (MiniCPM)","kind":"private","commercialActivityVerified":false,"shares":{"hosted-inference":{"low":0,"central":0.05,"high":0.3,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter listing at 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/api/v1/models","title":"OpenRouter model list (public API; models with a Hugging Face weights id and their per-token prices)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Entries with a hugging_face_id and a positive pricing.prompt; live listing, no publication date shown"},"note":"5LayerCake estimate (T5). Method: shares of tokens served through gateways, re-based to open-weight models only (a model counts as open when OpenRouter's model list gives it a Hugging Face weights id). Allowance only. No MiniCPM model is listed on OpenRouter. Gasgoo (2026-07-20) reports ModelBest's SuperMate cockpit product shipping in cars and KrASIA (2026-07-21) reports MiniCPM models announced for Samsung phones, which are on-device uses, not per-token sales; neither source confirms current paying use of the open MiniCPM weights. Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":1.2099,"central":1.2099,"high":1.2099,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. modelbest: 5.71M ÷ 472.00M = 1.210% (462 repositories). Largest: openbmb/MiniCPM5-2B 1.32M; openbmb/MiniCPM5-2B-DSpark 1.16M; cyankiwi/MiniCPM-SALA-AWQ-8bit 0.61M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 1.12%."},"local-device":{"low":0.5913,"central":0.6192,"high":0.6471,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. modelbest: 1.52M ÷ 234.28M = 0.647% (721 repositories). Largest: GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF 0.23M; GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking-GGUF 0.23M; openbmb/MiniCPM5-2B-GGUF 0.20M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. modelbest: 5.74M ÷ 970.10M = 0.591%. Central is the mean of the two channels (0.647 and 0.591), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"a1748b933e64a538a76c29d825bdd40811eed0a4c118ef04d679f154c22431f0"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"a1748b933e64a538a76c29d825bdd40811eed0a4c118ef04d679f154c22431f0"}}},"microsoft-phi":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"microsoft-phi","name":"Microsoft (Phi)","kind":"unit","parentId":"microsoft","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.0005,"central":0.0012,"high":0.5,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/microsoft","title":"OpenRouter author page for microsoft: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. microsoft-phi: 1.4B ÷ 113,640B = 0.001% (1 models; largest: microsoft/phi-4 1.4B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.001%. Range: the 60% band around the central, widened to the screening range (0.03 to 0.5%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.15% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 1 open-weight models of this developer, 1 with a positive per-token price (highest input price $0.07 per million tokens on microsoft/phi-4). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.7559,"central":0.7559,"high":0.7559,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. microsoft-phi: 3.57M ÷ 472.00M = 0.756% (3,490 repositories). Largest: microsoft/Phi-3.5-vision-instruct 0.64M; microsoft/phi-2 0.50M; microsoft/Phi-3.5-mini-instruct 0.38M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.70%."},"local-device":{"low":0.5321,"central":1.9371,"high":3.3422,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. microsoft-phi: 1.25M ÷ 234.28M = 0.532% (2,321 repositories). Largest: unsloth/Phi-4-mini-instruct-GGUF 0.11M; MaziyarPanahi/Phi-3.5-mini-instruct-GGUF 0.10M; MaziyarPanahi/phi-4-GGUF 0.10M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. microsoft-phi: 32.42M ÷ 970.10M = 3.342%. Central is the mean of the two channels (0.532 and 3.342), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":0.1096,"central":0.1096,"high":0.1096,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). microsoft-phi: 1.29B tokens ÷ 1,162.4B open-weight tokens = 0.111%, multiplied by 0.9914 as in the current estimate = 0.110% (6 models; largest: microsoft/phi-4 0.78B; microsoft/phi-4-reasoning-plus-04-30 0.44B; microsoft/phi-3.5-mini-128k-instruct 0.03B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":2.7808,"central":2.7808,"high":2.7808,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. microsoft-phi: 5.83M ÷ 209.81M = 2.781% (2,602 repositories). Largest: microsoft/Phi-3-mini-4k-instruct 1.41M; microsoft/phi-4 0.91M; microsoft/phi-2 0.65M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":2.4496,"central":3.4615,"high":4.4734,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. microsoft-phi: 0.82M ÷ 33.60M = 2.450% (2,064 repositories). Largest: MaziyarPanahi/Phi-3.5-mini-instruct-GGUF 0.12M; TheBloke/phi-2-GGUF 0.10M; MaziyarPanahi/Phi-4-mini-instruct-GGUF 0.09M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. microsoft-phi: 18.90M ÷ 422.60M = 4.473%. Central is the mean of the two channels (2.450 and 4.473), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":4814429,"central":4814429,"high":4814429,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 3.57M full-weight plus 1.25M GGUF and MLX downloads attributed to microsoft-phi in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":6657492,"central":6657492,"high":6657492,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 5.83M full-weight plus 0.82M GGUF and MLX downloads attributed to microsoft-phi in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c5c0ef7f741e4aac269ed5b44b984e91edb478723a12b746bc42d49a57c73d7e"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"c5c0ef7f741e4aac269ed5b44b984e91edb478723a12b746bc42d49a57c73d7e"}}},"poolside":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"poolside","name":"Poolside (Laguna)","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.05,"central":0.9214,"high":1.4742,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/poolside","title":"OpenRouter author page for poolside: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. poolside: 1,056.2B ÷ 113,640B = 0.929% (4 models; largest: poolside/laguna-s-2.1:free 943.0B; poolside/laguna-xs-2.1:free 80.5B; poolside/laguna-s-2.1 29.9B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.921%. Range: the 60% band around the central, widened to the screening range (0.05 to 1%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.3% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 4 open-weight models of this developer, 2 with a positive per-token price (highest input price $0.09 per million tokens on poolside/laguna-s-2.1). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.4044,"central":0.4044,"high":0.4044,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. poolside: 1.91M ÷ 472.00M = 0.404% (94 repositories). Largest: poolside/Laguna-XS.2 0.51M; poolside/Laguna-M.1 0.49M; poolside/Laguna-XS-2.1 0.43M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.37%."},"local-device":{"low":1.1315,"central":1.1315,"high":1.1315,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. poolside: 2.65M ÷ 234.28M = 1.131% (153 repositories). Largest: mudler/Laguna-XS-2.1-APEX-GGUF 1.65M; poolside/Laguna-S-2.1-GGUF 0.51M; poolside/Laguna-XS-2.1-GGUF 0.20M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"5c54d4c151d8398b6fc7cb8a9329097e1f4c7b2df5342a5bd43f009c08f7e19e"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"5c54d4c151d8398b6fc7cb8a9329097e1f4c7b2df5342a5bd43f009c08f7e19e"}}},"minimax":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"minimax","name":"MiniMax","kind":"public","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.4821,"central":1.2052,"high":8,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/minimax","title":"OpenRouter author page for minimax: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. minimax: 1,381.5B ÷ 113,640B = 1.216% (6 models; largest: minimax/minimax-m3 1,340.0B; minimax/minimax-m2.7 29.5B; minimax/minimax-m2.5 10.6B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 1.205%. Range: the 60% band around the central, widened to the screening range (0.5 to 8%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 2% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 6 open-weight models of this developer, 6 with a positive per-token price (highest input price $0.30 per million tokens on minimax/minimax-m3). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.4312,"central":0.4312,"high":0.4312,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. minimax: 2.04M ÷ 472.00M = 0.431% (174 repositories). Largest: MiniMaxAI/MiniMax-M2.5 0.78M; MiniMaxAI/MiniMax-M2.7 0.58M; MiniMaxAI/MiniMax-M3 0.20M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.40%."},"local-device":{"low":0.0747,"central":0.1902,"high":0.3057,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. minimax: 0.18M ÷ 234.28M = 0.075% (284 repositories). Largest: lmstudio-community/MiniMax-M2.5-MLX-8bit 0.02M; lmstudio-community/MiniMax-M2.5-MLX-4bit 0.01M; lmstudio-community/MiniMax-M2.5-MLX-6bit 0.01M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. minimax: 2.97M ÷ 970.10M = 0.306%. Central is the mean of the two channels (0.075 and 0.306), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":0.086,"central":0.086,"high":0.086,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). minimax: 1.01B tokens ÷ 1,162.4B open-weight tokens = 0.087%, multiplied by 0.9914 as in the current estimate = 0.086% (2 models; largest: minimax/minimax-01 0.81B; minimax/minimax-m1 0.20B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":0.0608,"central":0.0608,"high":0.0608,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. minimax: 0.13M ÷ 209.81M = 0.061% (8 repositories). Largest: MiniMaxAI/MiniMax-VL-01 0.10M; MiniMaxAI/MiniMax-M1-40k 0.02M; MiniMaxAI/MiniMax-Text-01-hf 0.01M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0,"central":0,"high":0,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. minimax: 0.00M ÷ 33.60M = 0.000% (0 repositories). Largest: none. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. minimax: 0.00M ÷ 422.60M = 0.000%. Central is the mean of the two channels (0.000 and 0.000), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope. Neither channel lists this developer at the time, so the observed share is 0 (a measured absence from complete listings, not a missing figure)."}},"activityGrowth":{"current":{"low":2210500,"central":2210500,"high":2210500,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 2.04M full-weight plus 0.18M GGUF and MLX downloads attributed to minimax in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":127610,"central":127610,"high":127610,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 0.13M full-weight plus 0.00M GGUF and MLX downloads attributed to minimax in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"baselineRevenueUsd":{"low":79038000,"central":79038000,"high":79038000,"unit":"USD","basis":"Total revenue, calendar year, US dollars","period":"CY2025","periodEnd":"2025-12-31","source":{"url":"https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results","title":"MiniMax Global Announces Full Year 2025 Financial Results","publisher":"MiniMax","publishedDate":"2026-03-02","accessedDate":"2026-10-22","tier":"T2","locator":"headline revenue US$79,038 thousand, up 158.9%; AI-native products US$53.1 million, Open Platform and enterprise services US$26.0 million"},"note":"The company says total revenue was US$79.0 million in 2025. Of this, US$53.1 million came from AI-native products and US$26.0 million from the Open Platform and enterprise services; the release does not say how much relates to open-weight models. MiniMax is the entity scored, so this is its own total revenue."},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"f52410750a51ba83d18c74a6ea5a742cc38498e8704f4c3b1d8b06b8bb55875d"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"f52410750a51ba83d18c74a6ea5a742cc38498e8704f4c3b1d8b06b8bb55875d"}}},"stepfun":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"stepfun","name":"StepFun","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.5,"central":1.4486,"high":2.5,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/stepfun","title":"OpenRouter author page for stepfun: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. stepfun: 1,660.4B ÷ 113,640B = 1.461% (3 models; largest: stepfun/step-5-preview 1,630.0B; stepfun/step-3.7-flash 29.1B; stepfun/step-3.5-flash 1.3B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 1.449%. Range: the 60% band around the central, widened to the screening range (0.5 to 2.5%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 1.5% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 3 open-weight models of this developer, 3 with a positive per-token price (highest input price $1.00 per million tokens on stepfun/step-5-preview). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.1947,"central":0.1947,"high":0.1947,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. stepfun: 0.92M ÷ 472.00M = 0.195% (54 repositories). Largest: stepfun-ai/GOT-OCR2_0 0.41M; stepfun-ai/Step-3.5-Flash 0.26M; stepfun-ai/Step-3.7-Flash-NVFP4 0.14M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.18%."},"local-device":{"low":0.0452,"central":0.0452,"high":0.0452,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. stepfun: 0.11M ÷ 234.28M = 0.045% (100 repositories). Largest: unsloth/Step-3.7-Flash-GGUF 0.07M; ggml-org/Step-3.5-Flash-GGUF 0.01M; stepfun-ai/Step-3.7-Flash-GGUF 0.01M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"1427a83b29983ac092c8cf9d85242df85f458995e66f2bd0bcc03c2f0bcb4e90"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"1427a83b29983ac092c8cf9d85242df85f458995e66f2bd0bcc03c2f0bcb4e90"}}},"thinking-machines":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"thinking-machines","name":"Thinking Machines (Inkling)","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.1,"central":0.3991,"high":2,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/thinkingmachines","title":"OpenRouter author page for thinkingmachines: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. thinking-machines: 457.5B ÷ 113,640B = 0.403% (4 models; largest: thinkingmachines/inkling:free 281.0B; thinkingmachines/inkling-small:free 146.0B; thinkingmachines/inkling-small 20.5B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.399%. Range: the 60% band around the central, widened to the screening range (0.1 to 2%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.6% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 4 open-weight models of this developer, 2 with a positive per-token price (highest input price $1.00 per million tokens on thinkingmachines/inkling). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.5498,"central":0.5498,"high":0.5498,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. thinking-machines: 2.60M ÷ 472.00M = 0.550% (105 repositories). Largest: thinkingmachines/Inkling 1.32M; thinkingmachines/Inkling-Small 0.77M; thinkingmachines/Inkling-Small-NVFP4 0.42M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.51%."},"local-device":{"low":0.5087,"central":0.5087,"high":0.5087,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. thinking-machines: 1.19M ÷ 234.28M = 0.509% (74 repositories). Largest: unsloth/Inkling-Small-GGUF 0.78M; unsloth/inkling-GGUF 0.40M; meshllm/inkling-UD-Q2_K_XL-layers 0.00M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"priorShares":{"hosted-inference":{"low":0,"central":0,"high":0,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). No model of thinking-machines appears among the 284 models in the week's data, so the observed share is 0 (a complete listing, not a missing figure). Inkling was published in July 2026 (OpenRouter model list creation date 2026-07-17). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":0,"central":0,"high":0,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. thinking-machines: 0.00M ÷ 209.81M = 0.000% (0 repositories). Largest: none. Not counted: ModelScope and other mirrors, which are not observed. No repository attributed to this developer exists in the snapshot, so the observed share is 0 (a measured absence from a full listing, not a missing figure). Its Inkling models were published from July 2026 (OpenRouter's model list gives the creation date 2026-07-17)."},"local-device":{"low":0,"central":0,"high":0,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. thinking-machines: 0.00M ÷ 33.60M = 0.000% (0 repositories). Largest: none. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope. No GGUF or MLX repository attributed to this developer exists in the snapshot, so the observed share is 0 (a measured absence from a full listing, not a missing figure)."}},"activityGrowth":{"current":{"low":3786899,"central":3786899,"high":3786899,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 2.60M full-weight plus 1.19M GGUF and MLX downloads attributed to thinking-machines in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":0,"central":0,"high":0,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 0.00M full-weight plus 0.00M GGUF and MLX downloads attributed to thinking-machines in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"9a5d68540c27af11a0e53bf810e514af7b786ba84a181c053f196a2528c2ff88"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"9a5d68540c27af11a0e53bf810e514af7b786ba84a181c053f196a2528c2ff88"}}},"ibm-granite":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"ibm-granite","name":"IBM Granite","kind":"unit","parentId":"ibm","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.0027,"central":0.0067,"high":0.6,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/ibm-granite","title":"OpenRouter author page for ibm-granite: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. ibm-granite: 7.7B ÷ 113,640B = 0.007% (2 models; largest: ibm-granite/granite-4.2-8b 6.8B; ibm-granite/granite-4.0-h-micro 0.9B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.007%. Range: the 60% band around the central, widened to the screening range (0.05 to 0.6%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.2% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 2 open-weight models of this developer, 2 with a positive per-token price (highest input price $0.06 per million tokens on ibm-granite/granite-4.2-8b). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.5604,"central":0.5604,"high":0.5604,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. ibm-granite: 2.65M ÷ 472.00M = 0.560% (960 repositories). Largest: ibm-granite/granite-4.1-3b 0.64M; ibm-granite/granite-3.3-2b-instruct 0.33M; ibm-granite/granite-docling-258M 0.29M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.52%."},"local-device":{"low":0.7778,"central":1.0535,"high":1.3292,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. ibm-granite: 1.82M ÷ 234.28M = 0.778% (1,564 repositories). Largest: ibm-granite/granite-4.2-8b-GGUF 0.26M; ibm-granite/granite-4.2-3b-GGUF 0.26M; ibm-granite/granite-4.2-30b-GGUF 0.25M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. ibm-granite: 12.89M ÷ 970.10M = 1.329%. Central is the mean of the two channels (0.778 and 1.329), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"priorShares":{"hosted-inference":{"low":0.0013,"central":0.0013,"high":0.0013,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). ibm-granite: 0.01B tokens ÷ 1,162.4B open-weight tokens = 0.001%, multiplied by 0.9914 as in the current estimate = 0.001% (1 models; largest: ibm-granite/granite-4.0-h-micro 0.01B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range."},"self-hosted":{"low":0.437,"central":0.437,"high":0.437,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. ibm-granite: 0.92M ÷ 209.81M = 0.437% (310 repositories). Largest: ibm-granite/granite-docling-258M 0.29M; ibm-granite/granite-3.3-8b-instruct 0.14M; ibm-granite/granite-3.3-2b-instruct 0.13M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0.621,"central":0.8099,"high":0.9988,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2025-10-15 (Hugging Face) and all time to 2025-10-14 (Ollama)","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. ibm-granite: 0.34M ÷ 33.60M = 0.999% (766 repositories). Largest: unsloth/granite-4.0-h-small-GGUF 0.03M; lmstudio-community/granite-4.0-h-tiny-GGUF 0.03M; ggml-org/granite-docling-258M-GGUF 0.02M. Ollama part: the library page https://web.archive.org/web/20251014055009/https://ollama.com/library (Internet Archive capture, 2025-10-14) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 422.60M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. ibm-granite: 2.62M ÷ 422.60M = 0.621%. Central is the mean of the two channels (0.999 and 0.621), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"activityGrowth":{"current":{"low":4467289,"central":4467289,"high":4467289,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 2.65M full-weight plus 1.82M GGUF and MLX downloads attributed to ibm-granite in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":1252480,"central":1252480,"high":1252480,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 0.92M full-weight plus 0.34M GGUF and MLX downloads attributed to ibm-granite in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"77d960408262eb20773918d18eaf523feca7c8e9e4afce014e878cc31d056960"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"77d960408262eb20773918d18eaf523feca7c8e9e4afce014e878cc31d056960"}}},"baidu-ernie":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"baidu-ernie","name":"Baidu ERNIE","kind":"unit","parentId":"baidu","commercialActivityVerified":false,"shares":{"hosted-inference":{"low":0.02,"central":0.2,"high":1,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter listing at 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/api/v1/models","title":"OpenRouter model list (public API; models with a Hugging Face weights id and their per-token prices)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Entries with a hugging_face_id and a positive pricing.prompt; live listing, no publication date shown"},"note":"5LayerCake estimate (T5). Method: shares of tokens served through gateways, re-based to open-weight models only (a model counts as open when OpenRouter's model list gives it a Hugging Face weights id). Allowance only. No ERNIE model appears among OpenRouter's open-weight listings and no hosted-volume figure for ERNIE 4.5 open weights was found. Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.5915,"central":0.5915,"high":0.5915,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. baidu-ernie: 2.79M ÷ 472.00M = 0.592% (128 repositories). Largest: baidu/Unlimited-OCR 1.48M; sahilchachra/Unlimited-OCR-AWQ 1.10M; baidu/ERNIE-4.5-21B-A3B-Thinking 0.09M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.55%."},"local-device":{"low":0.0559,"central":0.0559,"high":0.0559,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. baidu-ernie: 0.13M ÷ 234.28M = 0.056% (178 repositories). Largest: lmstudio-community/ERNIE-4.5-21B-A3B-MLX-4bit 0.02M; lmstudio-community/ERNIE-4.5-21B-A3B-MLX-6bit 0.02M; lmstudio-community/ERNIE-4.5-21B-A3B-MLX-8bit 0.02M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"priorShares":{"hosted-inference":{"low":0.0321,"central":0.0321,"high":0.0321,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter weekly bucket dated 2025-10-20","periodEnd":"2025-10-20","source":{"url":"https://web.archive.org/web/20251021131113/https://openrouter.ai/rankings","title":"OpenRouter LLM rankings page, Internet Archive capture of 21 October 2025 (embedded weekly usage of every listed model, bucket dated 20 October 2025)","publisher":"OpenRouter (via the Internet Archive)","publishedDate":"2025-10-21","accessedDate":"2026-10-22","tier":"T5","locator":"rankingData rows dated 2025-10-20 (prompt plus completion tokens per model); capture timestamp 20251021131113"},"note":"5LayerCake derivation (T5). Method: the capture embeds OpenRouter's usage for every model listed in the week dated 2025-10-20 (284 models; 4.606T prompt plus completion tokens, of which 1.162T (25.2%) on models whose weights are open). A model counts for the developer that trained its base weights (the same rule as for downloads: fine-tunes, merges and distills count for the base developer; closed API aliases and safety classifiers are left out). baidu-ernie: 0.38B tokens ÷ 1,162.4B open-weight tokens = 0.032%, multiplied by 0.9914 as in the current estimate = 0.032% (5 models; largest: baidu/ernie-4.5-300b-a47b 0.31B; baidu/ernie-4.5-21b-a3b-thinking 0.04B; baidu/ernie-4.5-vl-424b-a47b 0.01B). Comparability: the current estimate uses the same source (OpenRouter's token counts for every open-weight model, for the 7 days to 2026-10-22), so the two figures are on one basis; each is one week, so promotions and free periods are inside the range. No model of this developer is listed on OpenRouter now, so the current figure is an allowance, not a measurement."},"self-hosted":{"low":0.0605,"central":0.0605,"high":0.0605,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2025-10-15 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 209.81M. Left out as unattributable: 199,737 repositories with 25.04M downloads (10.7% of the 234.85M that the language-model repositories drew before this step), of which 125,251 repositories with 14.38M carried only a developer's architecture tag. baidu-ernie: 0.13M ÷ 209.81M = 0.061% (41 repositories). Largest: baidu/ERNIE-4.5-VL-28B-A3B-PT 0.08M; baidu/ERNIE-4.5-21B-A3B-PT 0.02M; baidu/ERNIE-4.5-0.3B-PT 0.01M. Not counted: ModelScope and other mirrors, which are not observed."},"local-device":{"low":0.5061,"central":0.5061,"high":0.5061,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2025-10-15. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 33.60M. Left out as unattributable: 77,537 repositories with 11.15M downloads (24.9% of the 44.75M that the language-model repositories drew before this step), of which 16,044 repositories with 0.96M carried only a developer's architecture tag. baidu-ernie: 0.17M ÷ 33.60M = 0.506% (93 repositories). Largest: lmstudio-community/ERNIE-4.5-21B-A3B-MLX-4bit 0.04M; lmstudio-community/ERNIE-4.5-21B-A3B-MLX-6bit 0.04M; lmstudio-community/ERNIE-4.5-21B-A3B-MLX-8bit 0.04M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"activityGrowth":{"current":{"low":2922925,"central":2922925,"high":2922925,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 2.79M full-weight plus 0.13M GGUF and MLX downloads attributed to baidu-ernie in the snapshot of 2026-10-14 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."},"prior":{"low":297047,"central":297047,"high":297047,"unit":"downloads","basis":"Downloads over the last 30 days, on Hugging Face, of the developer's language-model repositories (full-weight, GGUF and MLX), derivatives counted for the base-model developer by one rule; Hub weekly snapshots, same method in both periods","period":"30 days to 2025-10-15","periodEnd":"2025-10-15","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2025-10-15/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2025-10-15 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2025-10-15","accessedDate":"2026-10-22","tier":"T5","locator":"models/2025-10-15/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). 0.13M full-weight plus 0.17M GGUF and MLX downloads attributed to baidu-ernie in the snapshot of 2025-10-15 (rule in the share notes). Hugging Face only: ModelScope publishes no history, so it is not in either period."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"49389fa843ebcb638d3c44c1433b9bb6f99a3fcda6f808ee6262b5fce176d749"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"49389fa843ebcb638d3c44c1433b9bb6f99a3fcda6f808ee6262b5fce176d749"}}},"liquid-ai":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"liquid-ai","name":"Liquid AI","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.0079,"central":0.0196,"high":0.6,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/liquid","title":"OpenRouter author page for liquid: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. liquid-ai: 22.5B ÷ 113,640B = 0.020% (2 models; largest: liquid/lfm-2.5-2.6b:free 21.2B; liquid/lfm-2.5-24b-a2b 1.3B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.020%. Range: the 60% band around the central, widened to the screening range (0.03 to 0.6%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.15% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 2 open-weight models of this developer, 1 with a positive per-token price (highest input price $0.05 per million tokens on liquid/lfm-2.5-24b-a2b). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.2152,"central":0.2152,"high":0.2152,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. liquid-ai: 1.02M ÷ 472.00M = 0.215% (1,434 repositories). Largest: LiquidAI/LFM2.5-2.6B 0.12M; LiquidAI/LFM2.5-1.2B-Instruct 0.12M; LiquidAI/LFM2.5-230M 0.08M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.20%."},"local-device":{"low":0.2745,"central":0.9643,"high":1.654,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. liquid-ai: 3.88M ÷ 234.28M = 1.654% (1,753 repositories). Largest: LiquidAI/LFM2.5-2.6B-GGUF 0.84M; LiquidAI/LFM2.5-230M-GGUF 0.42M; LiquidAI/LFM2.5-1.2B-Instruct-GGUF 0.35M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. liquid-ai: 2.66M ÷ 970.10M = 0.275%. Central is the mean of the two channels (1.654 and 0.275), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"f496aca13d5d17b7998118883176ec9b114e61b13b7ec494785528402d479d40"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"f496aca13d5d17b7998118883176ec9b114e61b13b7ec494785528402d479d40"}}},"lg-ai-research":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"lg-ai-research","name":"LG AI Research (EXAONE)","kind":"unit","parentId":"lg","commercialActivityVerified":false,"shares":{"hosted-inference":{"low":0,"central":0.05,"high":0.3,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter listing at 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/api/v1/models","title":"OpenRouter model list (public API; models with a Hugging Face weights id and their per-token prices)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Entries with a hugging_face_id and a positive pricing.prompt; live listing, no publication date shown"},"note":"5LayerCake estimate (T5). Method: shares of tokens served through gateways, re-based to open-weight models only (a model counts as open when OpenRouter's model list gives it a Hugging Face weights id). Allowance only. No EXAONE model is listed on OpenRouter; EXAONE weights are on Hugging Face (1.16M 30-day downloads in all, mostly EXAONE 3.5 AWQ). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.2844,"central":0.2844,"high":0.2844,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. lg-ai-research: 1.34M ÷ 472.00M = 0.284% (254 repositories). Largest: LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-AWQ 0.78M; LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct 0.25M; LGAI-EXAONE/EXAONE-3.5-2.4B-Instruct 0.05M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.26%."},"local-device":{"low":0.0285,"central":0.0848,"high":0.141,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. lg-ai-research: 0.07M ÷ 234.28M = 0.029% (250 repositories). Largest: LGAI-EXAONE/EXAONE-4.0-1.2B-GGUF 0.00M; LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-GGUF 0.00M; LGAI-EXAONE/EXAONE-4.5-33B-GGUF 0.00M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. lg-ai-research: 1.37M ÷ 970.10M = 0.141%. Central is the mean of the two channels (0.029 and 0.141), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"26ecd5cf55e17c648cf430b58e4b78e77a22321c2c52bfc751af390e849ab1ee"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"26ecd5cf55e17c648cf430b58e4b78e77a22321c2c52bfc751af390e849ab1ee"}}},"bytedance-seed":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"bytedance-seed","name":"ByteDance Seed","kind":"unit","parentId":"bytedance","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.01,"central":0.1,"high":0.5,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"OpenRouter listing at 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/api/v1/models","title":"OpenRouter model list (public API; models with a Hugging Face weights id and their per-token prices)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"Entries with a hugging_face_id and a positive pricing.prompt; live listing, no publication date shown"},"note":"5LayerCake estimate (T5). Method: shares of tokens served through gateways, re-based to open-weight models only (a model counts as open when OpenRouter's model list gives it a Hugging Face weights id). Allowance only. UI-TARS 1.5 7B (ByteDance Seed) is priced on OpenRouter ($0.10 per million input tokens); the Seed-OSS language models have no listing found. Paid-hosting evidence: 1 listed models, 1 with positive prices (highest input price $0.10 per million tokens on bytedance/ui-tars-1.5-7b). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.1724,"central":0.1724,"high":0.1724,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. bytedance-seed: 0.81M ÷ 472.00M = 0.172% (273 repositories). Largest: LeaderboardModel1/zeta-2.1-autoround-W4A16 0.39M; ByteDance-Seed/UI-TARS-1.5-7B 0.24M; ByteDance-Seed/Seed-OSS-36B-Instruct 0.05M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.16%."},"local-device":{"low":0.1196,"central":0.1196,"high":0.1196,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. bytedance-seed: 0.28M ÷ 234.28M = 0.120% (294 repositories). Largest: mradermacher/zeta-2.1-GGUF 0.06M; lmstudio-community/Seed-OSS-36B-Instruct-MLX-8bit 0.02M; lmstudio-community/Seed-OSS-36B-Instruct-MLX-4bit 0.02M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"e3f7cdbe948360dd9023705677e017800715f0de2f83ff966041b9317a3d217d"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"e3f7cdbe948360dd9023705677e017800715f0de2f83ff966041b9317a3d217d"}}},"ant-group":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"ant-group","name":"Ant Group (Ling and Ring)","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.0468,"central":0.1169,"high":1.5,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/inclusionai","title":"OpenRouter author page for inclusionai: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. ant-group: 134.1B ÷ 113,640B = 0.118% (2 models; largest: inclusionai/ling-3.0-flash 132.0B; inclusionai/ling-3.0-flash-vl 2.0B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.117%. Range: the 60% band around the central, widened to the screening range (0.05 to 1.5%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.3% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 2 open-weight models of this developer, 2 with a positive per-token price (highest input price $0.06 per million tokens on inclusionai/ling-3.0-flash-vl). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.0843,"central":0.0843,"high":0.0843,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. ant-group: 0.40M ÷ 472.00M = 0.084% (147 repositories). Largest: inclusionAI/LLaDA2.0-mini 0.18M; inclusionAI/LLaDA2.1-mini 0.09M; inclusionAI/Ling-3.0-flash-VL 0.01M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.08%."},"local-device":{"low":0.4636,"central":0.4636,"high":0.4636,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. ant-group: 1.09M ÷ 234.28M = 0.464% (369 repositories). Largest: rapid-mlx/Ling-3.0-tiny-MLX-4bit 0.27M; bloomer010/Ling-3.0-tiny-GGUF 0.26M; AtomicChat/Ling-3.0-flash-GGUF 0.24M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"dc891c53fb247578f01fa9a8e5d15d025a5336d528708725824e98d5a2c015f4"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"dc891c53fb247578f01fa9a8e5d15d025a5336d528708725824e98d5a2c015f4"}}},"cohere":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"cohere","name":"Cohere","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.0541,"central":0.1353,"high":1.5,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/cohere","title":"OpenRouter author page for cohere: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. cohere: 155.0B ÷ 113,640B = 0.136% (2 models; largest: cohere/north-mini-code:free 155.0B; cohere/command-a 0.0B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.135%. Range: the 60% band around the central, widened to the screening range (0.1 to 1.5%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.4% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 2 open-weight models of this developer, 1 with a positive per-token price (highest input price $2.50 per million tokens on cohere/command-a). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.0556,"central":0.0556,"high":0.0556,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. cohere: 0.26M ÷ 472.00M = 0.056% (402 repositories). Largest: CohereLabs/c4ai-command-r-v01 0.07M; CohereLabs/North-Micro-Vision-Instruct 0.06M; CohereLabs/command-a-plus-05-2026-bf16 0.02M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.05%."},"local-device":{"low":0.1702,"central":0.352,"high":0.5337,"unit":"percent","basis":"Blend of two shares of the function's local-format use: Hugging Face downloads of GGUF and MLX repositories over the last 30 days, and Ollama library pulls over all time (cumulative; not the latest month)","offBasis":true,"period":"30 days to 2026-10-14 (Hugging Face) and all time to 2026-10-22 (Ollama)","periodEnd":"2026-10-22","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. cohere: 0.40M ÷ 234.28M = 0.170% (343 repositories). Largest: unsloth/North-Mini-Code-1.0-GGUF 0.23M; bartowski/c4ai-command-r7b-12-2024-GGUF 0.05M; DavidAU/Command-R-01-Ultra-NEO-V1-35B-IMATRIX-GGUF 0.01M. Ollama part: the library page https://ollama.com/library?sort=popular (opened 2026-10-22) shows cumulative all-time pulls per model, not a month; excluding embedding and safety-classifier models its pool is 970.10M pulls. Library names are attributed by the same base-model rule where the name shows the base (llama*, qwen*, gemma*, phi* followed by a digit, gpt-oss and similar; nemotron, the Llama-3.1-Nemotron-70B model, and phind-codellama, a CodeLlama fine-tune, count for Meta); deepseek-r1 (distills on Qwen and Llama bases) and fine-tunes that name no base stay unattributed; Ollama's library lists models, not repositories, so its pool keeps them. cohere: 5.18M ÷ 970.10M = 0.534%. Central is the mean of the two channels (0.170 and 0.534), which cover two local routes (Hugging Face downloads for llama.cpp and LM Studio; Ollama's own registry); low and high are the lower and higher of the two. The Ollama half is cumulative, not the latest month, so the observation is off-basis. Not counted: ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"26f2b3eca1b52c1bf095ac7fe9040b19c53b0028eb9b2983b0674f09e1b3e8da"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"26f2b3eca1b52c1bf095ac7fe9040b19c53b0028eb9b2983b0674f09e1b3e8da"}}},"reflection-ai":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"reflection-ai","name":"Reflection","kind":"private","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.01,"central":0.1108,"high":0.8,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/reflection","title":"OpenRouter author page for reflection: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. reflection-ai: 127.0B ÷ 113,640B = 0.112% (1 models; largest: reflection/beam-1.0-501b-a23b 127.0B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.111%. Range: the 60% band around the central, widened to the screening range (0.01 to 0.8%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.1% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 1 open-weight models of this developer, 1 with a positive per-token price (highest input price $0.25 per million tokens on reflection/beam-1.0-501b-a23b). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.0007,"central":0.0007,"high":0.0007,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. reflection-ai: 0.00M ÷ 472.00M = 0.001% (3 repositories). Largest: reflection/Beam-1.0-501B-A23B-NVFP4 0.00M; reflection/Beam-1.0-501B-A23B 0.00M; reflection/Beam-1.0-501B-A23B-FP8 0.00M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.00%."},"local-device":{"low":0.0004,"central":0.0004,"high":0.0004,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. reflection-ai: 0.00M ÷ 234.28M = 0.000% (3 repositories). Largest: unsloth/Beam-1.0-501B-A23B-GGUF 0.00M; AtomicChat/Beam-1.0-501B-A23B-GGUF 0.00M; DevQuasar/reflection.Beam-1.0-501B-A23B-GGUF 0.00M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"f27220b77f25cd86aa78b8fd651cdf7f41f3a2cc7f74cdf5ed83a0315824214d"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"f27220b77f25cd86aa78b8fd651cdf7f41f3a2cc7f74cdf5ed83a0315824214d"}}},"meituan":{"schema":"5lc-index-inputs/1","key":"models-open-weight","entityId":"meituan","name":"Meituan (LongCat)","kind":"public","commercialActivityVerified":true,"shares":{"hosted-inference":{"low":0.0035,"central":0.0087,"high":1.5,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"7 days to 2026-10-22","periodEnd":"2026-10-22","source":{"url":"https://openrouter.ai/meituan","title":"OpenRouter author page for meituan: models and tokens processed in the last 7 days (with the pages of the other 38 authors that have open-weight models)","publisher":"OpenRouter","publishedDate":"undated","accessedDate":"2026-10-22","tier":"T5","locator":"per-model 'Tokens processed in the last 7 days' on each author page; live pages, no publication date shown"},"note":"5LayerCake derivation (T5). Method: OpenRouter's author pages show the tokens each model processed in the last 7 days (prompt plus completion). I opened the pages of all 39 authors with a model that has a Hugging Face weights id in OpenRouter's public model list (opened 2026-10-22) and added up the tokens of the models that have one (free variants included), attributing each model to the developer of its base weights by the rule used for downloads (fine-tunes, merges and distills count for the base-model developer; closed models, embedding models and safety classifiers are left out). Open-weight tokens in the week: 113.64T across 176 models. meituan: 9.9B ÷ 113,640B = 0.009% (1 models; largest: meituan/longcat-2.0 9.9B), multiplied by 0.9914 to leave room for the allowances declared for the 14 candidates with no OpenRouter model = 0.009%. Range: the 60% band around the central, widened to the screening range (0.05 to 1.5%), which allows for hosting that OpenRouter does not see: the screening estimate this replaces was 0.3% (OpenRouter's week to 2026-10-21 scaled by 0.8, with Vercel's AI Gateway index for September 2026, which names only DeepSeek, Z.ai and Moonshot among open-weight developers, and the ATOM Report's January 2026 OpenRouter shares). Paid-hosting evidence: OpenRouter's public model list (https://openrouter.ai/api/v1/models, opened 2026-10-22) holds 1 open-weight models of this developer, 1 with a positive per-token price (highest input price $0.30 per million tokens on meituan/longcat-2.0). Gateway samples cover developer traffic, not large enterprise contracts or Chinese platforms (card known gaps)."},"self-hosted":{"low":0.0078,"central":0.0078,"high":0.0078,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: all model repositories in the Hugging Face weekly snapshot of 2026-10-14 (full hub, not a top-N list). One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable full-weight downloads over 30 days: 472.00M. Left out as unattributable: 270,693 repositories with 39.88M downloads (7.8% of the 511.88M that the language-model repositories drew before this step), of which 160,432 repositories with 23.58M carried only a developer's architecture tag. meituan: 0.04M ÷ 472.00M = 0.008% (28 repositories). Largest: meituan-longcat/LongCat-Flash-Chat 0.02M; meituan-longcat/LongCat-Flash-Lite-FP8 0.01M; meituan-longcat/LongCat-Flash-Lite 0.00M. Not counted: ModelScope and other mirrors, which are not observed. Including repositories created before 2023-07-18 the share would be 0.01%."},"local-device":{"low":0.0045,"central":0.0045,"high":0.0045,"unit":"percent","basis":"Share of the function's use of open-weight language models by base-model developer, latest available month: tokens served for hosted use, downloads of full weights for self-hosting, downloads and pulls of local-format copies for local use","period":"30 days to 2026-10-14","periodEnd":"2026-10-14","source":{"url":"https://huggingface.co/datasets/hfmlsoc/hub_weekly_snapshots/blob/main/models/2026-10-14/models.parquet","title":"Hugging Face Hub weekly snapshot of all models, 2026-10-14 (30-day downloads, tags, pipeline tag, creation date)","publisher":"Hugging Face (hfmlsoc, Hub weekly snapshots dataset)","publishedDate":"2026-10-14","accessedDate":"2026-10-22","tier":"T5","locator":"models/2026-10-14/models.parquet; columns id, downloads (last 30 days), tags, pipeline_tag, library_name, createdAt"},"note":"5LayerCake derivation (T5). Method: Hugging Face part from the weekly snapshot of 2026-10-14. One attribution rule for every developer and both dates: a repository counts for the developer that owns the organisation of the root repository reached by following its base_model tags (fine-tunes, quantised copies, merges and distills count for the base-model developer); failing that, for the developer that owns the publishing organisation; failing that, if it carries a base_model tag, for the developer whose family name the root repository's name carries (a declared copy or derivative whose chain ends outside any developer's organisation). A developer here is one of the scored companies or one of the other developers whose organisations the method lists (Allen AI, 01.AI, TII and similar). A repository with neither a chain nor a listed developer's organisation, or whose chain reaches no listed developer, is unattributable and is left out of the pool in both periods: a name or an architecture tag alone never credits a repository. NVIDIA repositories named Llama (the Llama-Nemotron models) and the copies derived from them count for Meta, as in the register scope and the OpenRouter mapping. Name matching uses whole family names: a name counts for StepFun only if it begins with Step and a digit, and for Moonshot only if Kimi is not followed by a letter (Kimiko fine-tunes of Llama do not count); TinyLlama and TinyMistral, separate models, count for neither Meta nor Mistral. Language-model repositories are those with pipeline tag text-generation, image-text-to-text or any-to-any (or an untagged GGUF, MLX or text-generation-inference repository). Left out: embedding, reranking and safety-classifier models (by pipeline tag, tag or name); repositories created before 2023-07-18 (GPT-2, OPT, Pythia and similar); test fixtures (the internal-testing organisations and nm-testing, repositories named tiny-random, unit-test, test-model or testing, and the tiny and story test models of ggml-org); third-party speculative-decoding draft and speculator models (names or tags such as DSpark, DFlash, Eagle3, speculator, draft-model), which are no developer's open-weight model, while a draft published by the developer's own organisation stays; and music, speech, video and image-generation models, which the card counts in Specialist models. An untagged repository is treated as one of those when it carries a media tag or library (music-generation, text-to-speech, text-to-video, diffusers, comfyui and similar), is named for a known media model or is a text-encoder or encoder-only copy, or when its base model's own pipeline tag is a media task; a repository with a language-model pipeline tag is kept. GGUF and MLX repositories count as local use, every other format (including quantised server formats) as self-hosted use. Pool of attributable GGUF and MLX downloads over 30 days: 234.28M. Left out as unattributable: 112,587 repositories with 66.01M downloads (22.0% of the 300.29M that the language-model repositories drew before this step), of which 27,405 repositories with 12.91M carried only a developer's architecture tag. meituan: 0.01M ÷ 234.28M = 0.005% (31 repositories). Largest: llmfan46/LongCat-Flash-Lite-Sparse-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-GGUF 0.00M; llmfan46/LongCat-Flash-Lite-Sparse-Ultra-Uncensored-Heretic-Native-MTP-And-LSA-Preserved-GGUF 0.00M; InquiringMinds-AI/LongCat-Flash-Lite-GGUF 0.00M. No model of this developer appears in Ollama's library, so the share is the Hugging Face part alone. Not counted: LM Studio and other local runners that do not publish counts, and ModelScope."}},"screening":{"hosted-inference":{"levelLow":1,"levelHigh":2},"self-hosted":{"levelLow":1,"levelHigh":2},"local-device":{"levelLow":1,"levelHigh":2}},"checks":{"sourceSupport":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"ce7fac63b5abbcacfb3cb15f6baef082fd5900a4b1a10e2d3e3482a0972bdce3"},"consistency":{"by":"Claude check pass, Sonnet 5.5","date":"2026-10-22","passed":true,"sha256":"ce7fac63b5abbcacfb3cb15f6baef082fd5900a4b1a10e2d3e3482a0972bdce3"}}}},"ratings":{"alibaba-qwen":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"alibaba-qwen","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Open-weight list below the top five: GLM-5.3-Flash 42, Muse Spark Open 39, Qwen3.8 27B 34, Beam 30, Nemotron 3.5 Super VL 26. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Qwen's listed open model, Qwen3.8 27B, scores 34, below at least six other open-weight models from five developers. Undated live page, opened 2026-10-22."}]}},"self-hosted":{"alternatives":{"level":2,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Tencent and Qwen 'cover the whole range' of sizes from under 1B upward (Sec. 1). Full date shown."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Google (82,506 derivatives) is the next-largest base family; Meta, NVIDIA, DeepSeek and Mistral follow (Sec. 1, 4). Full date shown."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Qwen 'has become part of the default workflow' for choosing models to fine-tune and deploy; 180-210 new derivatives a day (Sec. 4). Full date shown."},{"url":"https://a16z.com/ai-enterprise-2025/","date":"2025-06-10","support":"CIOs report rising switching costs; fine-tuned models carry high upfront costs, while prompts port more easily. Full date shown."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/articles/gemma-4-everything-you-need-to-know","date":"2026-04-06","support":"Gemma 4 31B (39) trails Qwen3.5 27B (42); Qwen3.5 35B A3B leads Gemma 4 26B A4B by 6 points. Full date shown."},{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"Qwen leads sub-10B downloads with 44%; DeepSeek holds 47% of the 250B+ segment (Fig. 12). arXiv v2 date."}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"GGUF (local-format) downloads a month: Qwen 39.6M, Gemma 20.8M, Llama 7.5M (Sec. 5). Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"'A model's adoption is largely determined within its first few months' after release (Sec. 2). Full date shown."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"GGUF (local-format) downloads a month: Qwen 39.6M, Gemma 20.8M, Llama 7.5M (Sec. 5). Full date shown."},{"url":"https://artificialanalysis.ai/articles/gemma-4-everything-you-need-to-know","date":"2026-04-06","support":"Gemma 4 31B (39) trails Qwen3.5 27B (42); Qwen3.5 35B A3B leads Gemma 4 26B A4B by 6 points. Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"alibaba-qwen","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Qwen3.8-Flash-Next (180B) and Qwen3.8 2.4T score 40 on index v4.3.2; seven other developers' open models score 41-46 (MiMo-V2.6-Pro, DeepSeek V4.1 Flash, GLM-5.3, Step 5 Preview, Kimi K3, GLM-5.3-Flash, Ling 3.1 Flash). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Open-weight use is already spread across labs: DeepSeek 58% of gateway tokens, Z.ai 7.2%, Moonshot 3.6%; Qwen3.8-Flash had 2.4% reach and 0.0% spend share in the launch scorecard. (Date form: full date.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Seven open models outscore Qwen3.8's 40 (best 46), so the best alternative does not fall short of Qwen here. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://openrouter.ai/api/v1/models/qwen/qwen3.8-27b/endpoints","date":"undated","support":"Qwen3.8 27B is already sold by 21 providers at $0.02-0.99 in / $0.50-3.20 out per million tokens, so hosted capacity is not tied to Alibaba. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"self-hosted":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Qwen3.8 27B scores 34; the next under-40B models are K2 Horizon MoVA 36B (25), AI9Stars 39B (22), Muse Glimmer 30B and Gemma 4 26B (17): no same-size rival at parity. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Larger rivals match Qwen3.8-Flash-Next (180B, score 40): GLM-5.3-Flash 320B (42), Ling 3.1 Flash 560B (41), Muse Spark Open 816B (39), MiMo-V2.6-Flash 309B (38), but need roughly 1.7-4.5x the memory. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://arxiv.org/html/2604.07190","date":"2026-05-25","support":"Qwen leads sub-10B downloads with 44% while DeepSeek leads 250B+ with 47% (Fig. 12), so other developers hold most downloads in every size band. (Date form: full date. Version 2 revised 25 May 2026; v1 was 2026-04-08.)"}]},"switchingTime":{"level":2,"evidence":[{"url":"https://arxiv.org/html/2604.07190","date":"2026-05-25","support":"Qwen is the base for 69% of new fine-tunes (Feb 2026) and has held 40%+ since May 2025 (section 5.1), so adapters and derivatives would need retraining. (Date form: full date. Version 2 revised 25 May 2026; v1 was 2026-04-08.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Company guide: leaving Qwen means new model ID plus new chat templates, sampling settings and tool-call parsing. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://www.constellationr.com/research/blog/evidence-shows-enterprises-use-open-weight-models-mainstream","date":"2026-09-26","support":"Analyst blog citing surveys: security and compliance review is the top obstacle (62%) for open-weight pilots; 46% of production users self-manage in cloud, 42% on premises. (Date form: full date.)"}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Qwen3.8 27B leads its class by 9 points (34 vs 25): a material gap, but only in the 27B tier. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"At the tiny end Qwen3.5 2B scores 7 against K2 Horizon 3.7B (16), MiniCPM5-2B (12), G9v3-3B (11) and Granite 4.2 3B (9): rivals lead there. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"From 180B up, rivals match or beat Qwen3.8 (40). Blending the tiers gives a small overall gap. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's popular list holds Llama 3.1 (120.5M pulls), Llama 3.2 (86M), Gemma 3 (41.3M), Mistral (34M), Gemma 4 (27.6M), Phi-3 (18.3M) and gpt-oss (14.1M) beside Qwen2.5 (43.2M), Qwen3 (41.2M) and Qwen3.5 (23.1M). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"Seven tiny open models score above Qwen3.5 2B (7): K2 Horizon 3.7B (16), MiniCPM5-2B (12), G9v3-3B (11), Granite 4.2 3B (9), MiniCPM5-1B (9), Nanbeige4.1-3B (8), LFM2.5-2.6B (8). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"At workstation size Qwen3.8 27B scores 34 against 17 for Gemma 4 26B A4B and Muse Glimmer 30B (and 15 for Gemma 4 31B): the best rival falls well short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"At phone size rivals lead Qwen3.5 2B (7), so the gap is real only for the larger local models; overall small. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"deepseek":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"deepseek","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."},{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"DeepSeek's OpenRouter open-model token share fell from 75.6% (June 2025) to 31.1% (January 2026) as Xiaomi reached 27.2% (Sec. 5.1, Fig. 9). arXiv v2 date."},{"url":"https://openrouter.ai/api/v1/models/deepseek/deepseek-v4.1-flash/endpoints","date":"undated","support":"V4.1 Flash weights are served by about 40 endpoints from some 30 providers, DeepSeek not among them, so hosting does not depend on it. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"DeepSeek had 58% of September open-weight tokens; V4.1-Flash priced about a fifth of the open-weight average. Full date shown."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"At the low-price 'flash' tier the nearest rival, GLM-5.3-Flash, scores 42 against V4.1 Flash's 45. Undated live page, opened 2026-10-22."}]}},"self-hosted":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Share of 2026 downloads above 70B: Moonshot 88%, DeepSeek 55%, Z.ai 39%, NVIDIA 14%, Meta 9%; MiniMax effectively all (Sec. 1). Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"'A model's adoption is largely determined within its first few months' after release (Sec. 2). Full date shown."},{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"Relative adoption at release: gpt-oss-120B about 19-20x, Nemotron Super 120B 11-17x, GLM-5 20x at 30 days (Table 1). arXiv v2 date."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"GGUF (local-format) downloads a month: Qwen 39.6M, Gemma 20.8M, Llama 7.5M (Sec. 5). Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"55% of DeepSeek's 2026 downloads are of models above 70B; GGUF leaders are Qwen, Gemma and Llama (Sec. 1, 5). Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"deepseek","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"DeepSeek V4.1 Flash scores 45; MiMo-V2.6-Pro 46, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44 and GLM-5.3-Flash 42 are within 3 points (five rivals). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Rivals already carry open-weight traffic at scale: Z.ai 7.2% and Moonshot 3.6% of gateway tokens beside DeepSeek's 58%. (Date form: full date.)"},{"url":"https://epoch.ai/data-insights/open-closed-eci-gap","date":"2026-05-29","support":"Epoch's May ranking of open models has three developers within 5 ECI points of the top (Kimi K2.6 151.6, MiniMax-M2.5 147.4, GLM-5 146.6). (Date form: full date.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"More than four in five DeepSeek tokens were served by DeepSeek itself, so rival hosts would have to add that capacity first. (Date form: full date.)"},{"url":"https://openrouter.ai/api/v1/models/deepseek/deepseek-v4.1-flash/endpoints","date":"undated","support":"32 providers already list DeepSeek V4.1 Flash on OpenRouter ($0.07-0.60 in, $0.18-2.40 out per million tokens), so the weights can be served elsewhere. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://api-docs.deepseek.com/quick_start/pricing","date":"undated","support":"DeepSeek lists V4.1 Flash at $0.30 in / $1.20 out per million tokens at peak and half that off-peak. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://docs.z.ai/guides/overview/pricing","date":"undated","support":"Z.ai lists GLM-5.3-Flash at $0.15 in / $0.50 out and GLM-5.3 at $1.40 / $4.40, so price-matched rivals exist. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"self-hosted":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"V4.1 Flash (552B, 16B active) scores 45; MiMo-V2.6-Pro 46, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44 and GLM-5.3-Flash 42 are within 3 points. The older V4 Pro 0813 scores 36, below ten other open models. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/deepseek-ai","date":"undated","support":"Most downloaded recent uploads: V4-Flash-0731 (4.3M), V4.1-Flash (1.49M), V4-Flash-Vision-Exp (386k), V4-Pro-0813 (152k); older versions have more rivals. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://arxiv.org/html/2604.07190","date":"2026-05-25","support":"DeepSeek holds 47% of downloads in the 250B+ band, so other developers hold the other 53% (Fig. 12). (Date form: full date. Version 2 revised 25 May 2026; v1 was 2026-04-08.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM's supported-models page lists about twenty open families (DeepSeek, Qwen, Gemma, Llama, Mistral, GLM, Kimi, MiniMax, Nemotron, gpt-oss, MiMo, Step, Hy, Granite) on one serving stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: only 11% changed model provider in a year, 66% upgraded within their provider, and switching to new models happens in weeks (all LLM users). (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"The best rival (MiMo-V2.6-Pro, 46) is ahead of DeepSeek V4.1 Flash (45), and rivals of 309B-2.8T parameters fit similar clusters. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM supports the other 250B+ families (GLM-5, Kimi-K3, MiniMax-M3, MiMo-V2 Pro, Step-5-Preview) on the same stack, so self-hosters keep their own hardware. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's deepseek-r1 (1.5b-671b, 94M pulls) sits beside Llama, Qwen, Gemma, Phi, Mistral and gpt-oss entries of 8M-120M pulls. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://itdaily.com/news/innovation/meta-muse-spark-open-weigh/","date":"2026-10-16","support":"Large local alternatives exist: Muse Spark Open (816B) is reported to run locally on two Mac Studios. (Date form: full date.)"},{"url":"https://huggingface.co/stepfun-ai","date":"undated","support":"StepFun publishes Step-3.7-Flash as a 197B GGUF for local runtimes, another large local rival. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Rivals at 180B-1T parameters score 38-46 against DeepSeek's 45 (V4.1 Flash) or 36 (V4 Pro 0813), so none falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Small local copies are widely replaced by newer Gemma 4 (27.6M pulls) and Qwen3.5 (23.1M) entries updated a week before the snapshot. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"google-deepmind":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"google-deepmind","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/articles/gemma-4-everything-you-need-to-know","date":"2026-04-06","support":"Gemma 4 31B (39) trails Qwen3.5 27B (42); Qwen3.5 35B A3B leads Gemma 4 26B A4B by 6 points. Full date shown."},{"url":"https://openrouter.ai/api/v1/models","date":"undated","support":"Hosted open models of Gemma 4's size include Qwen3.8-27B, Nemotron 3.5 Lightning 30B-A3B and Muse Glimmer 30B. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/articles/gemma-4-everything-you-need-to-know","date":"2026-04-06","support":"Gemma 4 31B (39) trails Qwen3.5 27B (42); Qwen3.5 35B A3B leads Gemma 4 26B A4B by 6 points. Full date shown."}]}},"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/articles/gemma-4-everything-you-need-to-know","date":"2026-04-06","support":"Gemma 4 31B (39) trails Qwen3.5 27B (42); Qwen3.5 35B A3B leads Gemma 4 26B A4B by 6 points. Full date shown."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Tencent and Qwen 'cover the whole range' of sizes from under 1B upward (Sec. 1). Full date shown."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Qwen has 151,448 derivatives on the Hub, 2.6 times Meta's and 4.7 times Llama's; Google has 82,506 (Sec. 4). Full date shown."},{"url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","date":"2026-04-02","support":"Google says Gemma has 400 million downloads and 'more than 100,000 variants'. Full date shown."},{"url":"https://a16z.com/ai-enterprise-2025/","date":"2025-06-10","support":"CIOs report rising switching costs; fine-tuned models carry high upfront costs, while prompts port more easily. Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/articles/gemma-4-everything-you-need-to-know","date":"2026-04-06","support":"Gemma 4 31B (39) trails Qwen3.5 27B (42); Qwen3.5 35B A3B leads Gemma 4 26B A4B by 6 points. Full date shown."}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."},{"url":"https://mistral.ai/news/mistral-3/","date":"2025-12-02","support":"Mistral says Ministral 3 (3B, 8B, 14B) targets edge and local use, all under Apache 2.0. Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","date":"2026-04-02","support":"Google says Gemma 4 comes in four sizes under Apache 2.0; E2B and E4B built with Pixel, Qualcomm and MediaTek. Full date shown."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"GGUF (local-format) downloads a month: Qwen 39.6M, Gemma 20.8M, Llama 7.5M (Sec. 5). Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"google-deepmind","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Gemma 4 26B A4B scores 17 and 31B scores 15; Qwen3.8 27B (34), K2 Horizon MoVA 36B (25), AI9Stars 39B (22) and Muse Glimmer 30B (17) are at or above. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/medium","date":"undated","support":"Mid-size open models from NVIDIA (26), Ant Group (25), Alibaba (16-18) and Mistral (14) also compete; API buyers can use any size. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Alibaba's Qwen3.8 27B scores 34 against Gemma 4's 15-17, so the best alternative is ahead, not behind. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Gemma 4 scores 15-17; four other developers' under-40B models score 17-34 (Alibaba, IFM, AI9Stars, Meta). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/medium","date":"undated","support":"Mid-size open models from NVIDIA (26), Ant Group (25) and Mistral (14) also score at or near Gemma 4, adding three more developers. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/google","date":"undated","support":"Gemma 4 31B-it and 26B-A4B-it show 8.63M and 9.44M downloads, last updated 20 Jul. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://arxiv.org/html/2604.07190","date":"2026-05-25","support":"Llama and Gemma each kept about 10% of new derivatives while Qwen held 40%+, so most fine-tuning already runs on other bases. (Date form: full date. Version 2 revised 25 May 2026; v1 was 2026-04-08.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 worked in vLLM, llama.cpp, MLX and LM Studio at launch; licence moved to Apache 2.0, the same as most rivals. (Date form: full date.)"},{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM's supported-models page lists about twenty open families (DeepSeek, Qwen, Gemma, Llama, Mistral, GLM, Kimi, MiniMax, Nemotron, gpt-oss, MiMo, Step, Hy, Granite) on one serving stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: only 11% changed model provider in a year, 66% upgraded within their provider, and switching to new models happens in weeks (all LLM users). (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Qwen3.8 27B (34) and K2 Horizon MoVA 36B (25) score above Gemma 4 31B (15): no rival falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Gemma 3 (41.3M pulls) and Gemma 4 (27.6M) compete with Llama, Qwen, Phi, Mistral and gpt-oss entries of 8M-120M pulls. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 E2B and E4B target phones, Raspberry Pi and Jetson, with 128K context. (Date form: full date.)"},{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"Rivals in the phone-size band include K2 Horizon 3.7B (16), MiniCPM5-2B (12), Granite 4.2 3B (9) and LFM2.5-2.6B (8). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"The twelve tiny open models listed score 6-16 and none is Gemma; AA_SMALL shows Gemma 4 at 15-17 against 34 for Qwen3.8 27B. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"At 26-31B the best rival (Qwen3.8 27B, 34) is far ahead of Gemma 4. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"zhipu":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"zhipu","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."},{"url":"https://openrouter.ai/api/v1/models/z-ai/glm-5.3/endpoints","date":"undated","support":"GLM-5.3 is served by 49 endpoints from many providers (Bedrock, Cloudflare, Fireworks), not only Z.ai. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Low-price flash-tier models: GLM-5.3-Flash scores 42, DeepSeek V4.1 Flash 45. Undated live page, opened 2026-10-22."}]}},"self-hosted":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Share of 2026 downloads above 70B: Moonshot 88%, DeepSeek 55%, Z.ai 39%, NVIDIA 14%, Meta 9%; MiniMax effectively all (Sec. 1). Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"'A model's adoption is largely determined within its first few months' after release (Sec. 2). Full date shown."},{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"Relative adoption at release: gpt-oss-120B about 19-20x, Nemotron Super 120B 11-17x, GLM-5 20x at 30 days (Table 1). arXiv v2 date."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"GGUF (local-format) downloads a month: Qwen 39.6M, Gemma 20.8M, Llama 7.5M (Sec. 5). Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"GGUF (local-format) downloads a month: Qwen 39.6M, Gemma 20.8M, Llama 7.5M (Sec. 5). Full date shown."},{"url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","date":"2026-04-02","support":"Google says Gemma 4 comes in four sizes under Apache 2.0; E2B and E4B built with Pixel, Qualcomm and MediaTek. Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"zhipu","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"GLM-5.3-Flash scores 42 (GLM-5.3 scores 45); MiMo-V2.6-Pro 46, DeepSeek 45, Step 5 Preview 44, Kimi K3 44, Ling 3.1 Flash 41 and both Qwen3.8 models (40) are within 3 points: seven rivals. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"GLM-5.3-Flash had 9.5% reach in the September launch scorecard; DeepSeek carries 58% of tokens and Moonshot 3.6%. (Date form: full date.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"MiMo-V2.6-Pro (46) and DeepSeek V4.1 Flash (45) outscore GLM-5.3-Flash (42) and match or beat GLM-5.3 (45). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://openrouter.ai/api/v1/models/z-ai/glm-5.3/endpoints","date":"undated","support":"38 providers list GLM-5.3 on OpenRouter, and US-based providers served seven in ten Z.ai tokens (Vercel), so hosted capacity is not tied to Zhipu. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://docs.z.ai/guides/overview/pricing","date":"undated","support":"Z.ai lists GLM-5.3-Flash at $0.15 in / $0.50 out and GLM-5.3 at $1.40 / $4.40. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://api-docs.deepseek.com/quick_start/pricing","date":"undated","support":"DeepSeek V4.1 Flash off-peak is $0.15 in / $0.60 out, so price-matched rivals exist. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"GLM-5.3-Flash (320B) scores 42; seven rivals score 40-46. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/zai-org","date":"undated","support":"GLM-5.3-Flash (321B) shows 6.42M downloads and GLM-5.3 (753B) 1.94M; Flash dominates, so Flash sets the comparison. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM's supported-models page lists about twenty open families (DeepSeek, Qwen, Gemma, Llama, Mistral, GLM, Kimi, MiniMax, Nemotron, gpt-oss, MiMo, Step, Hy, Granite) on one serving stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: only 11% changed model provider in a year, 66% upgraded within their provider, and switching to new models happens in weeks (all LLM users). (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"The best rival (46) is ahead of GLM-5.3-Flash (42) and GLM-5.3 (45). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's glm-ocr has 8.5M pulls (updated a week before the snapshot) among ~260 models. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM also supports OCR and vision-language models from other developers (DeepSeek-OCR, HunyuanOCR, Qwen3-VL, Gemma 4, MiniCPM-V, Kimi-VL, LFM2-VL). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://huggingface.co/zai-org","date":"undated","support":"GLM-OCR (1B) shows 1.68M downloads; no capability comparison against rival OCR/vision models was found, so evidence is thin and level 1 is the best-supported reading. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"meta-superintelligence-labs":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"meta-superintelligence-labs","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"Qwen reached 69% of new derivative models by February 2026; Llama fell to about 11%, Google about 10% (Sec. 5.1, Fig. 8). arXiv v2 date."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Qwen has 151,448 derivatives on the Hub, 2.6 times Meta's and 4.7 times Llama's; Google has 82,506 (Sec. 4). Full date shown."}]},"switchingTime":{"level":2,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Meta still has a large derivative footprint (Qwen's 151,448 is 2.6 times Meta's), so re-basing fine-tunes is work (Sec. 4). Full date shown."},{"url":"https://a16z.com/ai-enterprise-2025/","date":"2025-06-10","support":"CIOs report rising switching costs; fine-tuned models carry high upfront costs, while prompts port more easily. Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"After Llama 4, Meta 'collapsed on inference platforms and plateaued in adoption metrics' (Sec. 5.1 summary). arXiv v2 date."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Meta's newest open model, Muse Spark Open, scores 39, behind six open models from five other developers. Undated live page, opened 2026-10-22."}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"GGUF (local-format) downloads a month: Qwen 39.6M, Gemma 20.8M, Llama 7.5M (Sec. 5). Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"GGUF (local-format) downloads a month: Qwen 39.6M, Gemma 20.8M, Llama 7.5M (Sec. 5). Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/smollm3","date":"2025-07-08","support":"Hugging Face says SmolLM3 (3B) beats Llama-3.2-3B and Qwen2.5-3B, while Qwen3-4B generally scores highest. Full date shown."},{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"llama3.1 and llama3.2 copies were last updated one to two years ago, while gemma4 and qwen3.5 were updated last week. Undated live page, opened 2026-10-22."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"meta-superintelligence-labs","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Meta's Muse Glimmer 30B scores 17; Qwen3.8 27B (34), K2 Horizon MoVA 36B (25) and AI9Stars 39B (22) lead the under-40B class. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Muse Spark Open (816B, 42B active) scores 39, tenth of the twelve large open models shown (36-46). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://itdaily.com/news/innovation/meta-muse-spark-open-weigh/","date":"2026-10-16","support":"Muse Spark Open was released 15 Oct 2026 under the proprietary Muse Model License; weights on Hugging Face, SageMaker, Model Garden and Foundry. (Date form: full date.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://arxiv.org/html/2604.07190","date":"2026-05-25","support":"Meta's share of new fine-tunes peaked at 44% (Aug 2024) and fell to 11%; its OpenRouter token share went from 37.4% (Jan 2025) to zero within about a year (Aug 2025 in the summary, Jan 2026 in the Fig. 9 caption): users moved on. (Date form: full date. Version 2 revised 25 May 2026; v1 was 2026-04-08.)"},{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM supports Llama 2-4 alongside Qwen, Gemma, Mistral and others on one stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: switching to better new models happens in weeks. (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Rivals in Meta's own size class score 22-34 against Muse Glimmer's 17. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Ten open models outscore Muse Spark Open; none falls short of it. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Llama 3.1 (120.5M pulls, updated a year ago) and Llama 3.2 (86M, two years ago) face current Gemma 4 (27.6M), Qwen3.5 (23.1M) and gpt-oss (14.1M), all updated a week before the snapshot. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"No Llama appears among the twelve top tiny open models (scores 6-16). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://arxiv.org/html/2604.07190","date":"2026-05-25","support":"Meta's OpenRouter token share fell from 37.4% (Jan 2025) to zero within about a year as users moved to other open models (section 5.1). (Date form: full date. Version 2 revised 25 May 2026; v1 was 2026-04-08.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"Tiny rivals such as K2 Horizon 3.7B (16) and MiniCPM5-2B (12) are listed; no Llama is, so no rival falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"At 30B Meta's Muse Glimmer scores 17 against 34 for Qwen3.8 27B. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"nvidia-nemotron":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"nvidia-nemotron","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Open-weight list below the top five: GLM-5.3-Flash 42, Muse Spark Open 39, Qwen3.8 27B 34, Beam 30, Nemotron 3.5 Super VL 26. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://openrouter.ai/api/v1/models/nvidia/nemotron-3-ultra-550b-a55b:free/endpoints","date":"undated","support":"The free Nemotron 3 Ultra endpoint is served only by NVIDIA; paid copies run at DeepInfra, BaseTen and Venice (paid endpoint list). Undated live page, opened 2026-10-22."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/nvidia-nemotron-3-ultra-550b-a55b","date":"undated","support":"Nemotron 3 Ultra 550B A55B (Reasoning) scores 23 on Intelligence Index v4.3.2. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."}]}},"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"Relative adoption at release: gpt-oss-120B about 19-20x, Nemotron Super 120B 11-17x, GLM-5 20x at 30 days (Table 1). arXiv v2 date."},{"url":"https://openrouter.ai/api/v1/models","date":"undated","support":"Open models near Nemotron Super's 120B size include gpt-oss-120b, Qwen3.5-122B and Mistral Small 4 119B. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/","date":"2026-10-15","support":"NVIDIA says Super VL 'ships with open weights, open datasets and training recipes'; Bolt runs in llama.cpp, Ollama, vLLM. Full date shown."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"'A model's adoption is largely determined within its first few months' after release (Sec. 2). Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Nemotron 3.5 Super VL scores 26; ten listed open models from other developers score 30-46. Undated live page, opened 2026-10-22."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"nvidia-nemotron","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/medium","date":"undated","support":"Nemotron 3.5 Super VL (120B, 12B active) scores 26 and tops the mid-size class. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Large open models score 36-46 and Qwen3.8 27B 34: more than six rivals score higher. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Rivals score 36-46 against Nemotron's 26; none falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/medium","date":"undated","support":"Nemotron 3.5 Super VL (26) leads the mid-size class; Ling-3.0-flash VL scores 25 and Nemotron Bolt 21. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Qwen3.8 27B scores 34 and K2 Horizon MoVA 36B 25 at smaller size; large rivals score 36-46. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM's supported-models page lists about twenty open families (DeepSeek, Qwen, Gemma, Llama, Mistral, GLM, Kimi, MiniMax, Nemotron, gpt-oss, MiMo, Step, Hy, Granite) on one serving stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: only 11% changed model provider in a year, 66% upgraded within their provider, and switching to new models happens in weeks (all LLM users). (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Qwen3.8 27B (34) beats Nemotron 3.5 Super VL (26) at a quarter of the parameters, so no rival falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"tencent-hy":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"tencent-hy","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Open-weight list below the top five: GLM-5.3-Flash 42, Muse Spark Open 39, Qwen3.8 27B 34, Beam 30, Nemotron 3.5 Super VL 26. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."},{"url":"https://openrouter.ai/api/v1/models/tencent/hy4-preview/endpoints","date":"undated","support":"Hy4-preview is served by DeepInfra, Cohere, Novita and SiliconFlow as well as Tencent. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/hy3","date":"undated","support":"Hy3 scores 25 on Intelligence Index v4.3.2; no index score found for Hy4-preview, so evidence is thin. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"tencent-hy","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://huggingface.co/tencent","date":"undated","support":"Tencent's newest large model is Hy4-preview (780B, uploaded 28 Aug). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"No Tencent model is among the twelve top large open models (36-46) listed; rivals from eight developers fill the table. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Rivals score 36-46; Tencent's Hy4-preview is not listed, so the best alternative does not fall short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"xiaomi":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"xiaomi","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":2,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."},{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"DeepSeek's OpenRouter open-model token share fell from 75.6% (June 2025) to 31.1% (January 2026) as Xiaomi reached 27.2% (Sec. 5.1, Fig. 9). arXiv v2 date."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Flash-tier rivals to MiMo-V2.6-Flash: DeepSeek V4.1 Flash scores 45, GLM-5.3-Flash 42. Undated live page, opened 2026-10-22."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"xiaomi","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"MiMo-V2.6-Pro (1.02T, 42B active) tops the open ranking at 46; DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44 and Kimi K3 44 are within 3 points (four rivals). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In the September launch scorecard MiMo-V2.6-Pro reached 1.9% of teams and MiMo-V2.6-Flash 1.6%; DeepSeek, Z.ai and Moonshot carry far more open-weight tokens. (Date form: full date.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"},{"url":"https://arxiv.org/html/2604.07190","date":"2026-05-25","support":"Xiaomi went from zero to 27.2% of OpenRouter open-model tokens by Jan 2026 after MiMo-V2-Flash; Meta fell from 37.4% to zero: shares move in months. (Date form: full date. Version 2 revised 25 May 2026; v1 was 2026-04-08.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"The best rival, DeepSeek V4.1 Flash and GLM-5.3 at 45, is 1 point behind MiMo-V2.6-Pro (46): an effectively nil gap. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://openrouter.ai/api/v1/models/xiaomi/mimo-v2.6-pro/endpoints","date":"undated","support":"MiMo-V2.6-Pro has 5 providers at $0.39-0.52 in / $0.78-1.04 out; rival DeepSeek V4.1 Flash has 32 and GLM-5.3 has 38, so rival capacity is ample. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"openai":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"openai","rater":"a","model":"claude-opus-5-5","dachiReview":true,"ratedAt":"2026-10-22","functions":{"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"Relative adoption at release: gpt-oss-120B about 19-20x, Nemotron Super 120B 11-17x, GLM-5 20x at 30 days (Table 1). arXiv v2 date."},{"url":"https://openrouter.ai/api/v1/models","date":"undated","support":"Open models near gpt-oss-120b's size include Nemotron 3 Super 120B, Qwen3.5-122B and Mistral Small 4 119B. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"'A model's adoption is largely determined within its first few months' after release (Sec. 2). Full date shown."},{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"Relative adoption at release: gpt-oss-120B about 19-20x, Nemotron Super 120B 11-17x, GLM-5 20x at 30 days (Table 1). arXiv v2 date."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/gpt-oss-120b","date":"undated","support":"gpt-oss-120B (high) scores 12 on Intelligence Index v4.3.2. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Open-weight list below the top five: GLM-5.3-Flash 42, Muse Spark Open 39, Qwen3.8 27B 34, Beam 30, Nemotron 3.5 Super VL 26. Undated live page, opened 2026-10-22."}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","date":"2026-04-02","support":"Google says Gemma 4 comes in four sizes under Apache 2.0; E2B and E4B built with Pixel, Qualcomm and MediaTek. Full date shown."},{"url":"https://mistral.ai/news/mistral-3/","date":"2025-12-02","support":"Mistral says Ministral 3 (3B, 8B, 14B) targets edge and local use, all under Apache 2.0. Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"openai","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","dachiReview":true,"functions":{"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/medium","date":"undated","support":"gpt-oss-120b scores 12; NVIDIA's Nemotron 3.5 Super VL (26), Ant's Ling-3.0-flash VL (25), Qwen3.5 122B (16-18) and Mistral Medium 3.5 (14) score higher in the mid-size class. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Under-40B rivals from Alibaba, IFM, AI9Stars, Meta and Google score 15-34, also above gpt-oss-120b. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/openai","date":"undated","support":"gpt-oss-20b and 120b show 5.78M and 2.96M downloads, last updated 26 Aug 2025. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://arxiv.org/html/2604.07190","date":"2026-05-25","support":"New entrants win adoption fast: gpt-oss gathered more adoption than long-established organisations such as Mistral (section 5.1). (Date form: full date. Version 2 revised 25 May 2026; v1 was 2026-04-08.)"},{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM's supported-models page lists about twenty open families (DeepSeek, Qwen, Gemma, Llama, Mistral, GLM, Kimi, MiniMax, Nemotron, gpt-oss, MiMo, Step, Hy, Granite) on one serving stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: only 11% changed model provider in a year, 66% upgraded within their provider, and switching to new models happens in weeks (all LLM users). (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/medium","date":"undated","support":"Best mid-size rival (Nemotron 3.5 Super VL, 26) is more than twice gpt-oss-120b's 12; no rival falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"gpt-oss (20b, 120b) has 14.1M pulls against Gemma 3 41.3M, Gemma 4 27.6M, Qwen3.5 23.1M, Mistral 34M and Llama 3.1 120.5M. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Under-40B rivals score 15-34 on the index; gpt-oss-120b (the larger sibling) scores 12 and gpt-oss-20b is not on the list. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Listed under-40B rivals (Qwen3.8 27B 34, K2 Horizon MoVA 25, Gemma 4 and Muse Glimmer 15-17) are at least as strong as gpt-oss-120b (12); thin evidence for the 20b model. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"mistral":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"mistral","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://openrouter.ai/api/v1/models","date":"undated","support":"Mistral Small 4 (119B) competes with Nemotron 3 Super 120B, gpt-oss-120b and Qwen3.5-122B, all hosted openly. Undated live page, opened 2026-10-22."},{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"Qwen reached 69% of new derivative models by February 2026; Llama fell to about 11%, Google about 10% (Sec. 5.1, Fig. 8). arXiv v2 date."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"'A model's adoption is largely determined within its first few months' after release (Sec. 2). Full date shown."},{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"Relative adoption at release: gpt-oss-120B about 19-20x, Nemotron Super 120B 11-17x, GLM-5 20x at 30 days (Table 1). arXiv v2 date."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/mistral-small-4","date":"undated","support":"Mistral Small 4 (Reasoning) scores 11 on Intelligence Index v4.3.2. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Open-weight list below the top five: GLM-5.3-Flash 42, Muse Spark Open 39, Qwen3.8 27B 34, Beam 30, Nemotron 3.5 Super VL 26. Undated live page, opened 2026-10-22."}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://mistral.ai/news/mistral-3/","date":"2025-12-02","support":"Mistral says Ministral 3 (3B, 8B, 14B) targets edge and local use, all under Apache 2.0. Full date shown."},{"url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","date":"2026-04-02","support":"Google says Gemma 4 comes in four sizes under Apache 2.0; E2B and E4B built with Pixel, Qualcomm and MediaTek. Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"mistral","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/medium","date":"undated","support":"Mistral Medium 3.5 scores 14 and Mistral Small 4 scores 11; Nemotron 3.5 Super VL (26), Ling-3.0-flash VL (25), Ling-Fin (23), Nemotron Bolt (21) and Qwen3.5 122B (16-18) score higher. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/mistralai","date":"undated","support":"Mistral's newest large uploads are Mistral-Large-3 675B and Small-4 119B NVFP4 builds (Sept), plus Leanstral 1.5. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://arxiv.org/html/2604.07190","date":"2026-05-25","support":"Mistral led early (derivative share 57% in Dec 2023) but its downloads never matched Qwen or Llama: users moved on (section 5.1 summary). (Date form: full date. Version 2 revised 25 May 2026; v1 was 2026-04-08.)"},{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM's supported-models page lists about twenty open families (DeepSeek, Qwen, Gemma, Llama, Mistral, GLM, Kimi, MiniMax, Nemotron, gpt-oss, MiMo, Step, Hy, Granite) on one serving stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: only 11% changed model provider in a year, 66% upgraded within their provider, and switching to new models happens in weeks (all LLM users). (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/medium","date":"undated","support":"Best mid-size rival (26) is nearly double Mistral Medium 3.5 (14); no rival falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Mistral 7B has 34M pulls but was last updated a year ago; Gemma 4 (27.6M) and Qwen3.5 (23.1M) were updated a week before the snapshot. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"No Mistral model appears among the twelve tiny open models (scores 6-16). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"Twelve tiny rivals are ranked and none is Mistral; AA_SMALL puts Qwen3.8 27B at 34. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"moonshot":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"moonshot","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."},{"url":"https://openrouter.ai/api/v1/models/moonshotai/kimi-k3/endpoints","date":"undated","support":"Kimi K3 is served by 29 endpoints including Amazon Bedrock, Fireworks and Together, not only Moonshot. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."},{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Moonshot had 3.6% of tokens and 13.2% of September spend: valued, but rivals carry far more volume. Full date shown."}]}},"self-hosted":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Share of 2026 downloads above 70B: Moonshot 88%, DeepSeek 55%, Z.ai 39%, NVIDIA 14%, Meta 9%; MiniMax effectively all (Sec. 1). Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://techcrunch.com/2026/03/22/cursor-admits-its-new-coding-model-was-built-on-top-of-moonshot-ais-kimi/","date":"2026-03-22","support":"Cursor's Composer 2, launched in March 2026, was built on Kimi K2.5 with added training. Full date shown."},{"url":"https://techcrunch.com/2026/01/27/chinas-moonshot-releases-a-new-open-source-model-kimi-k2-5-and-a-coding-agent/","date":"2026-01-27","support":"Moonshot released Kimi K2.5 as an open model on 27 January, under two months before Composer 2. Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"moonshot","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Kimi K3 scores 44; MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, GLM-5.3-Flash 42 and Ling 3.1 Flash 41 are within 3 points (six rivals). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://epoch.ai/data-insights/open-closed-eci-gap","date":"2026-05-29","support":"Epoch's May list has MiniMax-M2.5 (147.4) and GLM-5 (146.6) within 5 ECI points of Kimi K2.6 (151.6). (Date form: full date.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":2,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"Kimi K3 accounted for nearly half of the increase in open-weight spend while processing far fewer tokens than DeepSeek V4.1 Flash; Moonshot had 3.6% of tokens but 13.2% of spend. (Date form: full date.)"},{"url":"https://openrouter.ai/api/v1/models/moonshotai/kimi-k3/endpoints","date":"undated","support":"Moonshot sells K3 at $3 in / $15 out per million tokens; 21 providers list it. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://api-docs.deepseek.com/quick_start/pricing","date":"undated","support":"DeepSeek lists V4.1 Flash, a 45-index rival, at $0.30 in / $1.20 out at peak, so buyers choosing K3 pay a premium. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Index gap is nil (rivals score 44-46), so the gap is small and rests on revealed willingness to pay. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Kimi K3 (2.8T, 104B active) scores 44; six rivals within 3 points are all smaller (MiMo-V2.6-Pro 1.02T 46, DeepSeek V4.1 Flash 552B 45, GLM-5.3 753B 45, Step 5 Preview 600B 44, GLM-5.3-Flash 320B 42, Ling 3.1 Flash 560B 41). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/moonshotai","date":"undated","support":"Kimi-K3 shows 955k downloads against 304k-361k for the K2.x models; the K3 weights are the largest listed. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM's supported-models page lists about twenty open families (DeepSeek, Qwen, Gemma, Llama, Mistral, GLM, Kimi, MiniMax, Nemotron, gpt-oss, MiMo, Step, Hy, Granite) on one serving stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: only 11% changed model provider in a year, 66% upgraded within their provider, and switching to new models happens in weeks (all LLM users). (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Rivals reach 46 (MiMo-V2.6-Pro) against K3's 44, so no rival falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://epoch.ai/data-insights/open-closed-eci-gap","date":"2026-05-29","support":"Epoch's open-weight ranking in May put Kimi K2.6 first, with MiniMax, Z.ai and Moonshot's other releases within about 5 ECI points. (Date form: full date.)"}]}}}}},"hugging-face":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"hugging-face","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/smollm3","date":"2025-07-08","support":"Hugging Face says SmolLM3 (3B) beats Llama-3.2-3B and Qwen2.5-3B, while Qwen3-4B generally scores highest. Full date shown."},{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"'A model's adoption is largely determined within its first few months' after release (Sec. 2). Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/smollm3","date":"2025-07-08","support":"Hugging Face says SmolLM3 (3B) beats Llama-3.2-3B and Qwen2.5-3B, while Qwen3-4B generally scores highest. Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"hugging-face","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://huggingface.co/HuggingFaceTB","date":"undated","support":"SmolLM3-3B (Sep 2025) shows 731k downloads and SmolLM2-135M-Instruct 1.19M; the newest uploads are small experiments (nanowhale-100m, May). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"The twelve tiny open models listed score 6-16 and none is SmolLM; K2 Horizon 3.7B (16), MiniCPM5-2B (12), G9v3-3B (11), Granite 4.2 3B (9) and others lead. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM lists SmolLM3 (via its Transformers backend) beside native Qwen, Gemma, Granite, LFM and MiniCPM entries, so a swap is a model-ID change. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM's supported-models page lists about twenty open families (DeepSeek, Qwen, Gemma, Llama, Mistral, GLM, Kimi, MiniMax, Nemotron, gpt-oss, MiMo, Step, Hy, Granite) on one serving stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: only 11% changed model provider in a year, 66% upgraded within their provider, and switching to new models happens in weeks (all LLM users). (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"At least eleven tiny rivals score 6-16; none falls short of SmolLM. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"modelbest":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"modelbest","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"self-hosted":{"alternatives":{"level":1,"evidence":[{"url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","date":"2026-04-02","support":"Google says Gemma 4 comes in four sizes under Apache 2.0; E2B and E4B built with Pixel, Qualcomm and MediaTek. Full date shown."},{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"Tencent and Qwen 'cover the whole range' of sizes from under 1B upward (Sec. 1). Full date shown."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://huggingface.co/blog/state-of-open-models-summer-2026","date":"2026-08-14","support":"'A model's adoption is largely determined within its first few months' after release (Sec. 2). Full date shown."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","date":"2026-04-02","support":"Google says all Gemma 4 models handle images and video natively; 'the E2B and E4B models feature native audio input'. Full date shown."},{"url":"https://huggingface.co/api/models/openbmb/MiniCPM-V-4_5","date":"undated","support":"MiniCPM-V 4.5 (Apache 2.0) drew about 86K downloads in 30 days, against about 4.0M for google/gemma-4-E4B-it. Undated live page, opened 2026-10-22."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"modelbest","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"self-hosted":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"MiniCPM5-2B scores 12, second of twelve; only K2 Horizon 3.7B (16) and G9v3-3B (11) are near, then Granite 4.2 3B (9). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"MiniCPM-V is a vision-language line; rivals served by vLLM include Qwen3-VL, Gemma 4, Phi-4-multimodal, Nemotron Nano VL, GLM-4.1V, Kimi-VL, LFM2-VL and SmolVLM2. No head-to-head comparison found, so thin. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/openbmb","date":"undated","support":"MiniCPM-V 4.7 35B-A3B and 1B variants show 138k and 78k downloads, updated days before the snapshot. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM lists MiniCPM, MiniCPM-V and MiniCPM-O beside the rival families, so swapping is a model-ID change plus retesting. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://docs.vllm.ai/en/latest/models/supported_models/","date":"undated","support":"vLLM's supported-models page lists about twenty open families (DeepSeek, Qwen, Gemma, Llama, Mistral, GLM, Kimi, MiniMax, Nemotron, gpt-oss, MiMo, Step, Hy, Granite) on one serving stack. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Mid-2025 survey: only 11% changed model provider in a year, 66% upgraded within their provider, and switching to new models happens in weeks (all LLM users). (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"K2 Horizon 3.7B (16) outscores MiniCPM5-2B (12), so the best rival does not fall short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"microsoft-phi":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"microsoft-phi","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","date":"2026-04-02","support":"Google says Gemma 4 comes in four sizes under Apache 2.0; E2B and E4B built with Pixel, Qualcomm and MediaTek. Full date shown."},{"url":"https://mistral.ai/news/mistral-3/","date":"2025-12-02","support":"Mistral says Ministral 3 (3B, 8B, 14B) targets edge and local use, all under Apache 2.0. Full date shown."},{"url":"https://huggingface.co/api/models/microsoft/Phi-4-mini-instruct","date":"undated","support":"Phi-4-mini-instruct (MIT, published February 2025) drew about 355K downloads in 30 days. Undated live page, opened 2026-10-22."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"microsoft-phi","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Phi-3 (18.3M pulls, updated two years ago) and Phi-4 (7.8M, a year ago) face current Gemma 4, Qwen3.5 and gpt-oss entries. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"Phi-4 Mini scores 6, last of the twelve tiny open models listed; eleven others score 6-16. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/microsoft","date":"undated","support":"Microsoft's recent Hugging Face uploads are Fara1.5, AesCode and MagenticBrain; no new Phi model is among them. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"Eleven tiny rivals score above or level with Phi-4 Mini (6); none falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"poolside":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"poolside","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://openrouter.ai/api/v1/models","date":"undated","support":"Open coding models hosted on OpenRouter include Qwen3-Coder-Next, Devstral 2, Kimi K2.7 Code, GLM-5.3 and DeepSeek V4. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://openrouter.ai/api/v1/models/poolside/laguna-s-2.1:free/endpoints","date":"undated","support":"The free Laguna S 2.1 endpoint is served only by Poolside. Undated live page, opened 2026-10-22."},{"url":"https://poolside.ai/models","date":"undated","support":"Poolside lists Laguna S 2.1 (118B) and XS 2.1 (33B) open weights; its API is 'free to use for a limited time'. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://poolside.ai/models","date":"undated","support":"The company says Laguna S 2.1 is 'competitive with models several times its size'; no independent index score found, so evidence is thin. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://poolside.ai/models","date":"undated","support":"Poolside lists Laguna S 2.1 (118B) and XS 2.1 (33B) open weights; its API is 'free to use for a limited time'. Undated live page, opened 2026-10-22."},{"url":"https://huggingface.co/api/models/poolside/Laguna-XS-2.1","date":"undated","support":"Laguna XS 2.1 (OpenMDW-1.1) drew about 478K downloads in 30 days, small next to Qwen's local volume. Undated live page, opened 2026-10-22."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"poolside","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://huggingface.co/poolside","date":"undated","support":"Laguna M.1 (226B), S 2.1 (118B) and XS 2.1 (33B) are Poolside's coding models; no Laguna model appears in Artificial Analysis's top-12 open tables (large 36-46, medium 11-26). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/moonshotai","date":"undated","support":"Coding rivals are open too: Moonshot's Kimi-K2.7-Code (1T, updated 15 Jun). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama lists qwen3-coder (30b, 480b; 10.2M pulls) and qwen2.5-coder (22.7M) in its popular models. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Poolside has no entry in Artificial Analysis's top-12 tables, where the measured large rivals score 36-46; no coding benchmark comparison was found, so evidence is thin and level 1 is the best-supported reading. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}},"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://huggingface.co/poolside","date":"undated","support":"Laguna-XS-2.1 (33B) shows 478k downloads and Laguna-tiny-per-element 171k; licences are OpenMDW or Apache 2.0. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Qwen3-coder (30b, 480b; 10.2M pulls) and qwen2.5-coder (0.5b-32b; 22.7M) cover the same local coding niche. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/small","date":"undated","support":"Under-40B rivals score 15-34 (Qwen3.8 27B 34); no Laguna model is listed, so thin evidence and level 1. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"minimax":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"minimax","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Open-weight list below the top five: GLM-5.3-Flash 42, Muse Spark Open 39, Qwen3.8 27B 34, Beam 30, Nemotron 3.5 Super VL 26. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."},{"url":"https://arxiv.org/abs/2604.07190","date":"2026-05-25","support":"MiniMax M2.1's relative adoption fell from 4.16x to 0.78x by 90 days as newer releases took over (Table 1). arXiv v2 date."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/minimax-m3","date":"undated","support":"MiniMax-M3 scores 29 on Intelligence Index v4.3.2. Undated live page, opened 2026-10-22."},{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"minimax","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/minimax-m3","date":"undated","support":"MiniMax-M3 (428B, 23B active) scores 29 on index v4.3.2 and is priced at $0.30 in / $1.20 out per million tokens. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"The twelve top large open models, from eight developers, all score 36-46, above M3's 29. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"All twelve listed large open models (36-46) outscore MiniMax-M3 (29). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/MiniMaxAI","date":"undated","support":"MiniMax's newest uploads are M3 (427B, 23 Jul) and M3.1-Flash (220B), so nothing newer changes the comparison. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"stepfun":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"stepfun","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":2,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two paid teams changed their most-used model, almost one in three their most-used lab (switching section). Full date shown."},{"url":"https://openrouter.ai/api/v1/models/stepfun/step-5-preview/endpoints","date":"undated","support":"Step 5 Preview is served by DeepInfra, Featherless and InferenceNet as well as StepFun. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source","date":"undated","support":"Intelligence Index v4.3.2 open-weight list: MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash 45, GLM-5.3 45, Step 5 Preview 44, Kimi K3 44. Undated live page, opened 2026-10-22."},{"url":"https://www.deeplearning.ai/the-batch/an-unexpected-open-weights-leader","date":"2026-10-02","support":"MiMo-V2.6-Pro scored 46, GLM-5.3 45 and Kimi K3 44 on the same index; MiMo weights under MIT. Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"stepfun","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"hosted-inference":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"Step 5 Preview (600B, 27B active) scores 44; MiMo-V2.6-Pro 46, DeepSeek 45, GLM-5.3 45, Kimi K3 44, GLM-5.3-Flash 42 and Ling 3.1 Flash 41 are within 3 points (six rivals). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/stepfun-ai","date":"undated","support":"Step-5-Preview (604B) was uploaded about 8 days before the snapshot; Step-3.7-Flash (201B, June) is the older model. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://vercel.com/blog/ai-gateway-production-index-october-2026","date":"2026-10-15","support":"In September nearly one in two gateway teams changed their most-used model; the report calls the model 'a replaceable part'. (Date form: full date.)"},{"url":"https://openrouter.ai/docs/quickstart","date":"undated","support":"OpenRouter's quickstart: set the 'model' field to any slug to switch; hundreds of models behind one endpoint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://menlovc.com/perspective/2025-mid-year-llm-market-update/","date":"2025-07-31","support":"Survey of builders (all LLMs, not only open): 'When new models come out, switching happens in weeks.' (Date form: full date.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/large","date":"undated","support":"The best rival (MiMo-V2.6-Pro, 46) is 2 points ahead of Step 5 Preview (44). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://openrouter.ai/api/v1/models/stepfun/step-5-preview/endpoints","date":"undated","support":"Step 5 Preview has 4 providers (InferenceNet, Featherless, DeepInfra, StepFun) at $1.00 in / $2.70 out; rivals have 5-38 providers, so capacity is not a constraint. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"thinking-machines":{},"ibm-granite":{"a":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"ibm-granite","rater":"a","model":"claude-opus-5-5","ratedAt":"2026-10-22","functions":{"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Local pulls: llama3.1 120.5M, deepseek-r1 94M, qwen2.5 43.2M, gemma3 41.3M, qwen3 41.2M, mistral 34M, phi3 18.3M, gpt-oss 14.1M. Undated live page, opened 2026-10-22."}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"One runtime library offers Meta, Alibaba, Google, Mistral, Microsoft, OpenAI and DeepSeek families side by side; switching is a new pull. Undated live page, opened 2026-10-22."}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://www.ibm.com/new/announcements/ibm-granite-4-0-hyper-efficient-high-performance-hybrid-models","date":"2025-10-02","support":"IBM says Granite 4.0 is Apache 2.0 and the first open family accredited under ISO/IEC 42001. Full date shown."},{"url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","date":"2026-04-02","support":"Google says Gemma 4 comes in four sizes under Apache 2.0; E2B and E4B built with Pixel, Qualcomm and MediaTek. Full date shown."}]}}}},"b":{"schema":"5lc-index-rating/1","key":"models-open-weight","entityId":"ibm-granite","rater":"b","model":"claude-sonnet-5-5","ratedAt":"2026-10-22","functions":{"local-device":{"alternatives":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"Granite 4.2 3B scores 9; K2 Horizon 3.7B (16), MiniCPM5-2B (12), G9v3-3B (11), MiniCPM5-1B (9), Nanbeige4.1-3B (8) and LFM2.5-2.6B (8) are within 3 points (six rivals). (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://huggingface.co/ibm-granite","date":"undated","support":"granite-4.2-3b and granite-4.2-8b show 40.5k and 86.1k downloads, updated two days before the snapshot. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"switchingTime":{"level":1,"evidence":[{"url":"https://ollama.com/library?sort=popular","date":"undated","support":"Ollama's library lists about 260 models that users pull by name; top entries come from Meta, Alibaba, Google, Microsoft, Mistral, OpenAI and DeepSeek. (Live page opened 2026-10-22; shows no date, so 'undated'.)"},{"url":"https://the-decoder.com/googles-gemma-4-is-now-available-with-apache-2-0-licensing-for-the-first-time/","date":"2026-04-02","support":"Gemma 4 launched on Hugging Face, Kaggle and Ollama and worked in vLLM, llama.cpp, MLX and LM Studio at launch, so a new local model is usable at once. (Date form: full date.)"},{"url":"https://docs.liquid.ai/guides/migration-guide","date":"undated","support":"Liquid AI's guide says moving from Qwen, Llama or Gemma is usually a model-ID and configuration change (templates, sampling and tool parsing to redo). (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]},"capabilityGap":{"level":1,"evidence":[{"url":"https://artificialanalysis.ai/models/open-source/tiny","date":"undated","support":"K2 Horizon 3.7B (16) and MiniCPM5-2B (12) outscore Granite 4.2 3B (9), so no rival falls short. (Live page opened 2026-10-22; shows no date, so 'undated'.)"}]}}}}},"baidu-ernie":{},"liquid-ai":{},"lg-ai-research":{},"bytedance-seed":{},"ant-group":{},"cohere":{},"reflection-ai":{},"meituan":{}},"validation":[]}