{ "name": "Bright Open Models comparison", "edition": "2026-10-11-r2", "publishedAt": "2026-10-11", "canonical": "https://brightaifuture.com/open-models", "citation": "Bright Open Models comparison, edition 2026-10-11-r2, https://brightaifuture.com/open-models/comparison/2026-10-11-r2.json", "license": { "name": "CC BY 4.0 for Bright original selection and annotations only", "url": "https://creativecommons.org/licenses/by/4.0/", "scope": "Bright grants CC BY 4.0 only for its original selection, arrangement and annotations in this comparison. Third-party model weights, code, datasets, quoted source material, trademarks and linked media retain their own rights and licenses. This grant does not license those materials." }, "unknownRule": "Not established means this review did not verify it, not that it is absent.", "methodology": "Fifteen named family records, nine artifact dimensions. States and findings are taken directly from the records. Source-check dates remain distinct from the edition publication date. No independent model reproduction, speed benchmark or hardware compatibility test was performed by Bright.", "dimensions": [ { "key": "weights", "label": "Weights", "description": "Downloadable learned parameters for local inspection, adaptation, or inference." }, { "key": "architecture", "label": "Architecture", "description": "A public description or configuration detailed enough to understand the model design." }, { "key": "inference", "label": "Inference code", "description": "Code or documented integrations for running the released weights." }, { "key": "trainingCode", "label": "Training code", "description": "Code used to pre-train or post-train the model, beyond inference and fine-tuning examples." }, { "key": "recipe", "label": "Training recipe", "description": "The stages, hyperparameters, mixtures, and other instructions needed to reproduce training." }, { "key": "dataInformation", "label": "Data information", "description": "A meaningful account of training-data composition, collection, filtering, and governance." }, { "key": "trainingData", "label": "Training data", "description": "The actual training corpus, or a reproducible path to reconstruct it." }, { "key": "evaluation", "label": "Evaluation", "description": "Published evaluation results, prompts, harnesses, or code that can be examined and repeated." }, { "key": "commercialUse", "label": "Commercial use", "description": "License terms that state whether and under what conditions commercial deployment is allowed." } ], "rows": [ { "id": "qwen", "name": "Qwen3.8", "organization": "Qwen / Alibaba Cloud", "version": "Qwen3.8-27B", "releaseDate": "2026-08-14", "dateEvents": [ { "kind": "availability", "date": "2026-08-14", "precision": "day", "label": "Qwen3.8-27B availability", "sourceUrl": "https://github.com/QwenLM/Qwen3.8" } ], "summary": "A downloadable 27B vision-language member of the Qwen3.8 generation, selected here because its Apache-2.0 terms are materially simpler than the custom license on the largest Qwen3.8 checkpoint.", "parameters": "27B", "totalBillions": 27, "activeBillions": 27, "architecture": "Dense multimodal causal transformer; official configuration and implementation are published with the weights.", "modalities": ["text", "image"], "context": "262K tokens in the official deployment guidance", "license": "Apache-2.0", "commercialUse": "Permitted by Apache-2.0, subject to its notice and attribution terms.", "openness": [ { "key": "weights", "status": "available", "detail": "Post-trained weights and configuration are downloadable.", "sourceUrl": "https://huggingface.co/Qwen/Qwen3.8-27B" }, { "key": "architecture", "status": "available", "detail": "Configuration and model implementation are public.", "sourceUrl": "https://huggingface.co/Qwen/Qwen3.8-27B" }, { "key": "inference", "status": "available", "detail": "The official repository documents local deployment.", "sourceUrl": "https://github.com/QwenLM/Qwen3.8" }, { "key": "trainingCode", "status": "not-established", "detail": "The reviewed release does not establish publication of the complete training code.", "sourceUrl": "https://huggingface.co/Qwen/Qwen3.8-27B" }, { "key": "recipe", "status": "partial", "detail": "Technical materials describe model construction, but do not provide a fully reproducible end-to-end recipe.", "sourceUrl": "https://huggingface.co/Qwen/Qwen3.8-27B" }, { "key": "dataInformation", "status": "partial", "detail": "High-level training information is published; the complete mixture is not disclosed.", "sourceUrl": "https://huggingface.co/Qwen/Qwen3.8-27B" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://huggingface.co/Qwen/Qwen3.8-27B" }, { "key": "evaluation", "status": "partial", "detail": "The model card reports vendor evaluations; complete independent reproduction is not bundled.", "sourceUrl": "https://huggingface.co/Qwen/Qwen3.8-27B" }, { "key": "commercialUse", "status": "available", "detail": "This 27B checkpoint is Apache-2.0; other Qwen3.8 sizes must be checked separately.", "sourceUrl": "https://huggingface.co/Qwen/Qwen3.8-27B" } ], "sources": [ { "id": "qwen38-repo", "title": "Qwen3.8 official repository", "url": "https://github.com/QwenLM/Qwen3.8", "type": "repository" }, { "id": "qwen38-card", "title": "Qwen3.8-27B model card", "url": "https://huggingface.co/Qwen/Qwen3.8-27B", "type": "registry" } ], "strengths": [ "Multimodal input", "Long context", "Permissive license on this checkpoint" ], "limitations": [ "Vendor benchmark claims need independent reproduction", "Licensing differs across Qwen3.8 sizes" ], "quantizations": [ "Community quantizations exist; compatibility and provenance vary" ], "memory": "No single minimum is asserted; memory depends on precision, runtime, and context length.", "discoveryIds": ["qwen3"], "independentEvidence": "No current independent evaluation is attached to this family record; displayed capability claims should be read as vendor-reported.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/qwen", "recordUrl": "https://brightaifuture.com/open-models/qwen", "licenseSummary": "Commercial use permitted under Apache-2.0; review accompanying terms.", "licenseUrl": "https://huggingface.co/Qwen/Qwen3.8-27B", "download": { "gb": 30.866866928, "label": "About 30.87 GB of selected documented weights", "checkpoint": "Qwen/Qwen3.8-27B-FP8", "format": "Safetensors, official FP8", "sourceUrl": "https://huggingface.co/Qwen/Qwen3.8-27B-FP8/tree/017b9c7af6b5689d5dd426a76e0bc077eb5ca20a", "checkedAt": "2026-10-11", "scope": "Official FP8 variant of Qwen3.8-27B. All 64 layers-* files, mtp.safetensors and outside.safetensors; unique weight-index set. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 30866866928, "revision": "017b9c7af6b5689d5dd426a76e0bc077eb5ca20a", "fileCount": 66, "metadataUrl": "https://huggingface.co/api/models/Qwen/Qwen3.8-27B-FP8/revision/017b9c7af6b5689d5dd426a76e0bc077eb5ca20a?blobs=true", "modelCardUrl": "https://huggingface.co/Qwen/Qwen3.8-27B-FP8/blob/017b9c7af6b5689d5dd426a76e0bc077eb5ca20a/README.md", "licenseUrl": "https://huggingface.co/Qwen/Qwen3.8-27B-FP8/blob/017b9c7af6b5689d5dd426a76e0bc077eb5ca20a/LICENSE", "artifactLicense": "apache-2.0", "publisher": "maker" } }, { "id": "deepseek", "name": "DeepSeek V4.1 Flash", "organization": "DeepSeek", "version": "DeepSeek-V4.1-Flash", "releaseDate": "2026-09-10", "dateEvents": [ { "kind": "availability", "date": "2026-09-10", "precision": "day", "label": "DeepSeek API availability and new pricing", "sourceUrl": "https://api-docs.deepseek.com/news/news260910/" } ], "releaseNotes": "The cited event establishes API availability and pricing on this date; it does not independently timestamp the weight upload.", "summary": "A downloadable DeepSeek multimodal Flash model, with sparse activation that differs between prompt processing and generation.", "parameters": "552B backbone; 8B active in prefill / 16B in decode. Hugging Face displays 763B in model files; no single total-storage estimate is asserted.", "totalBillions": null, "activeBillions": 16, "architecture": "Sparse mixture-of-experts multimodal transformer; 8B active during prefill and 16B during decoding.", "modalities": ["text", "image"], "context": "1M tokens", "license": "MIT", "commercialUse": "Permitted by the MIT license.", "openness": [ { "key": "weights", "status": "available", "detail": "Official weights are downloadable.", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" }, { "key": "architecture", "status": "available", "detail": "The model card documents the sparse architecture and activation behavior.", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" }, { "key": "inference", "status": "available", "detail": "Official prompt encoding and inference examples are provided.", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" }, { "key": "trainingCode", "status": "not-established", "detail": "Complete pre-training and post-training code was not confirmed.", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" }, { "key": "recipe", "status": "partial", "detail": "The release explains major techniques but is not an end-to-end reproducible recipe.", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" }, { "key": "dataInformation", "status": "partial", "detail": "Some training approach information is disclosed without a complete corpus inventory.", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" }, { "key": "evaluation", "status": "available", "detail": "Evaluation reproduction instructions are included with the release.", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" }, { "key": "commercialUse", "status": "available", "detail": "The official checkpoint is MIT-licensed.", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" } ], "sources": [ { "id": "deepseek-v41-announcement", "title": "DeepSeek V4.1 Flash announcement", "url": "https://api-docs.deepseek.com/news/news260910/", "type": "institution" }, { "id": "deepseek-v41-card", "title": "DeepSeek-V4.1-Flash model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash", "type": "registry" } ], "strengths": ["Long context", "Multimodal input", "Sparse inference"], "limitations": [ "Displayed file size and stated backbone parameter count differ", "Training corpus is not released" ], "quantizations": [ "Official and community formats should be verified against the model card before use" ], "memory": "No reliable universal minimum is stated; the very large checkpoint requires distributed or aggressively quantized serving.", "discoveryIds": [], "independentEvidence": "No independent evaluation is attached to this record.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/deepseek", "recordUrl": "https://brightaifuture.com/open-models/deepseek", "licenseSummary": "Commercial use permitted under MIT; retain the required notice.", "licenseUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash", "download": { "gb": 510.296708312, "label": "About 510.30 GB of selected documented weights", "checkpoint": "deepseek-ai/DeepSeek-V4.1-Flash", "format": "Safetensors, mixed FP8 / FP4 expert weights", "sourceUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/tree/2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "checkedAt": "2026-10-11", "scope": "Full checkpoint, not an estimate from active parameters. Does not resolve 552B-backbone versus 763B registry accounting. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 510296708312, "revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "fileCount": 48, "metadataUrl": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4.1-Flash/revision/2cba9e42aa026125f3ed06c6d98c1db82f7ca027?blobs=true", "modelCardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/2cba9e42aa026125f3ed06c6d98c1db82f7ca027/README.md", "licenseUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/2cba9e42aa026125f3ed06c6d98c1db82f7ca027/LICENSE", "artifactLicense": "mit", "publisher": "maker" } }, { "id": "mistral", "name": "Mistral Small 4", "organization": "Mistral AI", "version": "Mistral-Small-4-119B-2603", "releaseDate": "2026-03-16", "dateEvents": [ { "kind": "initial-release", "date": "2026-03-16", "precision": "day", "label": "Mistral Small 4 release", "sourceUrl": "https://docs.mistral.ai/models/mistral-small-4-0-26-03" } ], "summary": "A hybrid sparse model that combines instruction following, reasoning, visual understanding, and coding modes in one downloadable checkpoint.", "parameters": "119B total / 6.5B active in Mistral’s model docs; its release article says 6B active per token (8B including embedding and output layers)", "totalBillions": 119, "activeBillions": 6.5, "architecture": "Mixture-of-experts transformer with 128 experts and four active experts per token.", "modalities": ["text", "image"], "context": "256K tokens", "license": "Apache-2.0", "commercialUse": "Permitted by Apache-2.0.", "openness": [ { "key": "weights", "status": "available", "detail": "Official FP8 weights are downloadable.", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }, { "key": "architecture", "status": "available", "detail": "Architecture, expert counts, and configuration are published.", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }, { "key": "inference", "status": "available", "detail": "The card documents vLLM and Transformers serving.", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }, { "key": "trainingCode", "status": "not-established", "detail": "The full code used to train the model was not confirmed.", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }, { "key": "recipe", "status": "partial", "detail": "The release describes capabilities and design without a full reproducible recipe.", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }, { "key": "dataInformation", "status": "not-established", "detail": "A sufficiently detailed training-data account was not confirmed.", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }, { "key": "evaluation", "status": "partial", "detail": "Vendor results are published, including machine-readable evaluation artifacts.", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }, { "key": "commercialUse", "status": "available", "detail": "Apache-2.0 permits commercial use.", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" } ], "sources": [ { "id": "mistral-small4-release", "title": "Introducing Mistral Small 4", "url": "https://mistral.ai/news/mistral-small-4/", "type": "institution" }, { "id": "mistral-small4-card", "title": "Mistral Small 4 model card", "url": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603", "type": "registry" } ], "strengths": [ "Reasoning mode", "Visual understanding", "Agentic coding", "Efficient active parameter count" ], "limitations": [ "Training data is not documented in enough detail for reconstruction", "Performance claims are primarily vendor-reported" ], "quantizations": ["FP8", "Official NVFP4 checkpoint"], "memory": "The FP8 repository is about 242 GB; the NVFP4 repository is about 71 GB before runtime overhead.", "discoveryIds": [], "independentEvidence": "The registry records vendor evaluations but does not treat them as independent corroboration.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/mistral", "recordUrl": "https://brightaifuture.com/open-models/mistral", "licenseSummary": "Commercial use permitted under Apache-2.0; review accompanying terms.", "licenseUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603", "download": { "gb": 70.808028368, "label": "About 70.81 GB of selected documented weights", "checkpoint": "mistralai/Mistral-Small-4-119B-2603-NVFP4", "format": "Safetensors, official NVFP4", "sourceUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603-NVFP4/tree/45331841b631f4e281df8e959ea3cc9beb84298a", "checkedAt": "2026-10-11", "scope": "Official NVFP4 variant, all 13 consolidated shards. Excludes separate FP8 repository and alternate representations. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 70808028368, "revision": "45331841b631f4e281df8e959ea3cc9beb84298a", "fileCount": 13, "metadataUrl": "https://huggingface.co/api/models/mistralai/Mistral-Small-4-119B-2603-NVFP4/revision/45331841b631f4e281df8e959ea3cc9beb84298a?blobs=true", "modelCardUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603-NVFP4/blob/45331841b631f4e281df8e959ea3cc9beb84298a/README.md", "licenseUrl": "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603-NVFP4/blob/45331841b631f4e281df8e959ea3cc9beb84298a/README.md", "artifactLicense": "apache-2.0", "publisher": "maker" } }, { "id": "gemma", "name": "Gemma 4", "organization": "Google DeepMind", "version": "Gemma 4 26B-A4B", "releaseDate": "2026-03-31", "dateEvents": [ { "kind": "initial-release", "date": "2026-03-31", "precision": "day", "label": "Initial Gemma 4 model release", "sourceUrl": "https://ai.google.dev/gemma/docs/releases" }, { "kind": "announcement", "date": "2026-04-02", "precision": "day", "label": "Google Gemma 4 announcement", "sourceUrl": "https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/" } ], "releaseNotes": "The 26B-A4B reference checkpoint was in the initial March 31 release; later Gemma 4 updates are separate events.", "summary": "Google’s Gemma 4 generation spans edge and workstation sizes; this atlas record uses the 26B mixture-of-experts checkpoint as its reference point.", "parameters": "26B total / 4B active (reference checkpoint)", "totalBillions": 26, "activeBillions": 4, "architecture": "Hybrid local/global-attention multimodal transformer; the 26B reference variant uses mixture-of-experts routing.", "modalities": ["text", "image"], "context": "Up to 256K tokens on the 26B reference variant", "license": "Apache-2.0", "commercialUse": "Permitted by Apache-2.0.", "openness": [ { "key": "weights", "status": "available", "detail": "Pre-trained and instruction-tuned weights are released.", "sourceUrl": "https://ai.google.dev/gemma/docs/core/model_card_4" }, { "key": "architecture", "status": "available", "detail": "The official card describes the dense and MoE variants and attention design.", "sourceUrl": "https://ai.google.dev/gemma/docs/core/model_card_4" }, { "key": "inference", "status": "available", "detail": "Google documents supported runtimes and deployment routes.", "sourceUrl": "https://ai.google.dev/gemma/docs/core/model_card_4" }, { "key": "trainingCode", "status": "not-established", "detail": "The complete code used for the original training runs was not confirmed.", "sourceUrl": "https://ai.google.dev/gemma/docs/core/model_card_4" }, { "key": "recipe", "status": "partial", "detail": "The model card documents methods but does not provide a full reproduction recipe.", "sourceUrl": "https://ai.google.dev/gemma/docs/core/model_card_4" }, { "key": "dataInformation", "status": "partial", "detail": "Data categories and filtering are described at a high level.", "sourceUrl": "https://ai.google.dev/gemma/docs/core/model_card_4" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://ai.google.dev/gemma/docs/core/model_card_4" }, { "key": "evaluation", "status": "partial", "detail": "Official benchmark methodology and results are published; full independent reproduction is separate.", "sourceUrl": "https://ai.google.dev/gemma/docs/core/model_card_4" }, { "key": "commercialUse", "status": "available", "detail": "Gemma 4 is released under Apache-2.0.", "sourceUrl": "https://ai.google.dev/gemma/docs/core/model_card_4" } ], "sources": [ { "id": "gemma4-release-log", "title": "Gemma release log", "url": "https://ai.google.dev/gemma/docs/releases", "type": "institution" }, { "id": "gemma4-release", "title": "Gemma 4: Byte for byte, the most capable open models", "url": "https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/", "type": "institution" }, { "id": "gemma4-card", "title": "Gemma 4 model card", "url": "https://ai.google.dev/gemma/docs/core/model_card_4", "type": "report" } ], "strengths": [ "Reasoning and tool use", "Vision input", "Wide size range", "Official QAT checkpoints" ], "limitations": [ "Audio is limited to the E2B, E4B, and 12B variants", "Training data is described but not released" ], "quantizations": ["Official quantization-aware-trained checkpoints"], "memory": "Google documents a 12B variant running in 16 GB; requirements for the 26B-A4B variant depend on format and context.", "discoveryIds": [], "independentEvidence": "Arena placement cited by Google is a third-party leaderboard result, but this record does not independently reproduce it.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/gemma", "recordUrl": "https://brightaifuture.com/open-models/gemma", "licenseSummary": "Commercial use permitted under Apache-2.0; review accompanying terms.", "licenseUrl": "https://ai.google.dev/gemma/docs/core/model_card_4", "download": { "gb": 15.634191744, "label": "About 15.63 GB of selected documented weights", "checkpoint": "google/gemma-4-26B-A4B-it-qat-q4_0-gguf", "format": "GGUF Q4_0 QAT plus multimodal projector", "sourceUrl": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf/tree/d1c082be9cf3c8a514acf63b8761f4b41935842e", "checkedAt": "2026-10-11", "scope": "Google-published QAT variant, not original BF16 checkpoint. 14,439,363,584-byte language GGUF plus 1,194,828,160-byte projector. Distinct artifact; does not recheck original evaluation or nine-door findings. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 15634191744, "revision": "d1c082be9cf3c8a514acf63b8761f4b41935842e", "fileCount": 2, "metadataUrl": "https://huggingface.co/api/models/google/gemma-4-26B-A4B-it-qat-q4_0-gguf/revision/d1c082be9cf3c8a514acf63b8761f4b41935842e?blobs=true", "modelCardUrl": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf/blob/d1c082be9cf3c8a514acf63b8761f4b41935842e/README.md", "licenseUrl": "https://ai.google.dev/gemma/docs/gemma_4_license", "artifactLicense": "apache-2.0", "publisher": "maker" } }, { "id": "gpt-oss", "name": "gpt-oss", "organization": "OpenAI", "version": "gpt-oss-120b and gpt-oss-20b", "releaseDate": "2025-08-05", "dateEvents": [ { "kind": "initial-release", "date": "2025-08-05", "precision": "day", "label": "gpt-oss-120b and gpt-oss-20b release", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" } ], "summary": "Two open-weight text reasoning models with adjustable reasoning effort, structured output, and tool-use support.", "parameters": "117B / 5.1B active; 21B / 3.6B active", "totalBillions": 117, "activeBillions": 5.1, "architecture": "Transformer mixture-of-experts with alternating dense and locally banded sparse attention.", "modalities": ["text"], "context": "128K tokens", "license": "Apache-2.0, with a separate usage policy", "commercialUse": "Apache-2.0 permits commercial use; deployers should also review the published usage policy.", "openness": [ { "key": "weights", "status": "available", "detail": "Both checkpoints are downloadable.", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" }, { "key": "architecture", "status": "available", "detail": "The model card publishes architecture and parameter details.", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" }, { "key": "inference", "status": "available", "detail": "Reference PyTorch and Metal implementations and Harmony renderers are public.", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" }, { "key": "trainingCode", "status": "not-established", "detail": "The complete training stack was not released.", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" }, { "key": "recipe", "status": "partial", "detail": "Pre- and post-training approaches are described without a fully reproducible recipe.", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" }, { "key": "dataInformation", "status": "partial", "detail": "OpenAI says the dataset is mostly English and text-only, focused on STEM, code, and general knowledge.", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" }, { "key": "evaluation", "status": "partial", "detail": "A detailed model card and vendor evaluations are published.", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" }, { "key": "commercialUse", "status": "available", "detail": "Apache-2.0 permits broad commercial use.", "sourceUrl": "https://openai.com/index/introducing-gpt-oss/" } ], "sources": [ { "id": "gptoss-release", "title": "Introducing gpt-oss", "url": "https://openai.com/index/introducing-gpt-oss/", "type": "institution" }, { "id": "gptoss-card", "title": "gpt-oss model card", "url": "https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf", "type": "report" } ], "strengths": [ "Tool use", "Adjustable reasoning effort", "Local deployment", "Structured outputs" ], "limitations": [ "Text-only", "Training corpus is not released", "Self-hosters own system-level safeguards" ], "quantizations": ["Official MXFP4"], "memory": "OpenAI states that 120b fits in 80 GB and 20b in 16 GB in the released MXFP4 format.", "discoveryIds": ["gpt-oss"], "independentEvidence": "OpenAI publishes extensive internal evaluations; independent reproductions are not incorporated in this record.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/gpt-oss", "recordUrl": "https://brightaifuture.com/open-models/gpt-oss", "licenseSummary": "Commercial use permitted under Apache-2.0; review accompanying terms.", "licenseUrl": "https://openai.com/index/introducing-gpt-oss/", "download": { "publisher": "maker", "gb": 13.76, "label": "About 13.76 GB of listed model shards", "checkpoint": "openai/gpt-oss-20b", "format": "Safetensors with MXFP4 expert weights", "sourceUrl": "https://huggingface.co/openai/gpt-oss-20b/tree/main", "checkedAt": "2026-10-11", "scope": "Smaller member named in this family record. Sum of the three rounded shard listings, excluding supporting files. Other artifact formats differ; this is not runtime memory." } }, { "id": "llama", "name": "Llama 4", "organization": "Meta", "version": "Scout and Maverick", "releaseDate": "2025-04-05", "dateEvents": [ { "kind": "initial-release", "date": "2025-04-05", "precision": "day", "label": "Llama 4 Scout and Maverick release", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" } ], "summary": "Meta’s first natively multimodal Llama generation and first Llama mixture-of-experts release.", "parameters": "Scout: 109B / 17B active; Maverick: 400B / 17B active", "totalBillions": 400, "activeBillions": 17, "architecture": "Early-fusion multimodal mixture-of-experts transformer.", "modalities": ["text", "image"], "context": "Scout: 10M tokens; Maverick: 1M tokens", "license": "Llama 4 Community License Agreement", "commercialUse": "Commercial use is allowed subject to the custom license, including an additional term for services above 700 million monthly active users.", "openness": [ { "key": "weights", "status": "available", "detail": "Scout and Maverick weights are available under Meta’s custom terms.", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "key": "architecture", "status": "available", "detail": "Meta publishes architecture and model-card details.", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "key": "inference", "status": "available", "detail": "Official downloads and ecosystem integrations support local inference.", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "key": "trainingCode", "status": "not-established", "detail": "The complete code used to train Llama 4 was not released.", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "key": "recipe", "status": "partial", "detail": "Meta describes key methods including multimodal pre-training and distillation.", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "key": "dataInformation", "status": "partial", "detail": "The model card provides categories and governance information without the corpus.", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "key": "evaluation", "status": "partial", "detail": "Meta reports benchmark results; full evaluation reproduction is not bundled.", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "key": "commercialUse", "status": "restricted", "detail": "Commercial use is governed by a custom community license with scale-based conditions.", "sourceUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" } ], "sources": [ { "id": "llama4-release", "title": "The Llama 4 herd", "url": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/", "type": "institution" } ], "strengths": [ "Native multimodality", "Very long context on Scout", "Large deployment ecosystem" ], "limitations": [ "Custom license is not OSI-approved", "Training data and full training stack are not released" ], "quantizations": [ "Meta states Scout can fit on one H100 with Int4 quantization" ], "memory": "Scout is described as fitting on one H100 in Int4; Maverick is described as fitting on one H100 host.", "discoveryIds": [], "independentEvidence": "This record does not independently reproduce Meta’s comparative benchmark claims.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/llama", "recordUrl": "https://brightaifuture.com/open-models/llama", "licenseSummary": "Custom or artifact-specific terms: read them. Commercial use is allowed subject to the custom license, including an additional term for services above 700 million monthly active users.", "licenseUrl": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/", "download": { "gb": 217.28373872, "label": "About 217.28 GB of selected documented weights", "checkpoint": "meta-llama/Llama-4-Scout-17B-16E-Instruct", "format": "Safetensors, BF16", "sourceUrl": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct/tree/92f3b1597a195b523d8d9e5700e57e4fbb8f20d3", "checkedAt": "2026-10-11", "scope": "Scout Instruct only, 109B total and 17B active, not Maverick. Public API lists 50 shards. Gated index/config returned 401; no login/agreement acceptance or bypass. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 217283738720, "revision": "92f3b1597a195b523d8d9e5700e57e4fbb8f20d3", "fileCount": 50, "metadataUrl": "https://huggingface.co/api/models/meta-llama/Llama-4-Scout-17B-16E-Instruct/revision/92f3b1597a195b523d8d9e5700e57e4fbb8f20d3?blobs=true", "modelCardUrl": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct/blob/92f3b1597a195b523d8d9e5700e57e4fbb8f20d3/README.md", "licenseUrl": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct/blob/92f3b1597a195b523d8d9e5700e57e4fbb8f20d3/LICENSE", "artifactLicense": "Llama 4 Community License Agreement", "publisher": "maker" } }, { "id": "olmo", "name": "Olmo 3", "organization": "Allen Institute for AI (Ai2)", "version": "Olmo 3 7B and 32B", "releaseDate": "2025-11-20", "dateEvents": [ { "kind": "initial-release", "date": "2025-11-20", "precision": "day", "label": "Olmo 3 model-flow release", "sourceUrl": "https://allenai.org/blog/olmo3" }, { "kind": "update", "date": "2025-12-12", "precision": "day", "label": "Olmo 3.1 checkpoint update", "sourceUrl": "https://allenai.org/blog/olmo3" } ], "releaseNotes": "This record begins with the Olmo 3 flow and its cited source also documents the later Olmo 3.1 checkpoint update.", "summary": "A fully open model flow spanning base, mid-trained, long-context, instruction, reasoning, and RL-zero checkpoints.", "parameters": "7B and 32B", "totalBillions": 32, "activeBillions": 32, "architecture": "Dense decoder-only transformer with a documented multi-stage training pipeline.", "modalities": ["text"], "context": "Up to 65K tokens in the long-context models", "license": "Apache-2.0 for models and code; data components retain their stated licenses", "commercialUse": "Model and code licenses permit commercial use; downstream users must review dataset licenses.", "openness": [ { "key": "weights", "status": "available", "detail": "Final and intermediate checkpoints are released.", "sourceUrl": "https://allenai.org/blog/olmo3" }, { "key": "architecture", "status": "available", "detail": "Architecture and changes are documented.", "sourceUrl": "https://allenai.org/blog/olmo3" }, { "key": "inference", "status": "available", "detail": "Model code and deployment instructions are public.", "sourceUrl": "https://allenai.org/blog/olmo3" }, { "key": "trainingCode", "status": "available", "detail": "Pre-training and post-training code are released.", "sourceUrl": "https://allenai.org/blog/olmo3" }, { "key": "recipe", "status": "available", "detail": "Training stages, configurations, checkpoints, and mixtures are documented.", "sourceUrl": "https://allenai.org/blog/olmo3" }, { "key": "dataInformation", "status": "available", "detail": "Ai2 publishes detailed provenance and composition for Dolma 3 and post-training data.", "sourceUrl": "https://allenai.org/blog/olmo3" }, { "key": "trainingData", "status": "available", "detail": "Dolma 3 and the released post-training datasets are accessible.", "sourceUrl": "https://allenai.org/blog/olmo3" }, { "key": "evaluation", "status": "available", "detail": "Evaluation suites, data, and scoring standards are public.", "sourceUrl": "https://allenai.org/blog/olmo3" }, { "key": "commercialUse", "status": "available", "detail": "Apache-2.0 covers model artifacts; constituent data licenses still apply.", "sourceUrl": "https://allenai.org/blog/olmo3" } ], "sources": [ { "id": "olmo3-release", "title": "Olmo 3: Charting a path through the model flow", "url": "https://allenai.org/blog/olmo3", "type": "institution" }, { "id": "olmo3-home", "title": "Olmo model flow", "url": "https://allenai.org/olmo", "type": "institution" } ], "strengths": [ "End-to-end inspectability", "Intermediate checkpoints", "Open data", "Reasoning research" ], "limitations": [ "Text-only", "Full reproduction still requires substantial compute" ], "quantizations": [ "Community quantizations are available; no single format is treated as canonical here" ], "memory": "No universal minimum is asserted; 7B and 32B variants span consumer to multi-GPU deployments.", "discoveryIds": ["olmo-2-32b", "goodfire-olmo-post-training"], "independentEvidence": "Ai2 documents external research using the open stack, including Goodfire’s post-training behavior tracing; those case studies do not validate every capability claim.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/olmo", "recordUrl": "https://brightaifuture.com/open-models/olmo", "licenseSummary": "Commercial use permitted under Apache-2.0; review accompanying terms.", "licenseUrl": "https://allenai.org/blog/olmo3", "download": { "gb": 14.596063712, "label": "About 14.60 GB of selected documented weights", "checkpoint": "allenai/Olmo-3-7B-Instruct", "format": "Safetensors, BF16", "sourceUrl": "https://huggingface.co/allenai/Olmo-3-7B-Instruct/tree/6e5971d9eba42665f5bd5a0fcf047f299ce1dccc", "checkedAt": "2026-10-11", "scope": "7B Instruct only, not 32B, Think or Olmo 3.1. Three shards. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 14596063712, "revision": "6e5971d9eba42665f5bd5a0fcf047f299ce1dccc", "fileCount": 3, "metadataUrl": "https://huggingface.co/api/models/allenai/Olmo-3-7B-Instruct/revision/6e5971d9eba42665f5bd5a0fcf047f299ce1dccc?blobs=true", "modelCardUrl": "https://huggingface.co/allenai/Olmo-3-7B-Instruct/blob/6e5971d9eba42665f5bd5a0fcf047f299ce1dccc/README.md", "licenseUrl": "https://huggingface.co/allenai/Olmo-3-7B-Instruct/blob/6e5971d9eba42665f5bd5a0fcf047f299ce1dccc/README.md", "artifactLicense": "apache-2.0", "publisher": "maker" } }, { "id": "nemotron", "name": "Nemotron 3 Nano Omni", "organization": "NVIDIA", "version": "Nemotron-3-Nano-Omni-30B-A3B-Reasoning", "releaseDate": "2026-04-28", "dateEvents": [ { "kind": "announcement", "date": "2026-04-27", "precision": "day", "label": "NVIDIA technical report publication", "sourceUrl": "https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Omni-report.pdf" }, { "kind": "availability", "date": "2026-04-28", "precision": "day", "label": "NVIDIA Build availability", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" } ], "summary": "A compact sparse reasoning model that accepts video, audio, images, and text while activating about three billion parameters per token.", "parameters": "33B listed deployment parameters on NVIDIA Build; the same page’s architecture text says 31B A3B, while the model identifier names a 30B-A3B backbone. Accounting differs across first-party fields.", "totalBillions": 33, "activeBillions": 3, "architecture": "Native omni-understanding sparse mixture-of-experts model in the Nemotron 3 family.", "modalities": ["text", "image", "audio", "video"], "context": "262K tokens on NVIDIA Build; the April 27 technical report describes a 256K maximum context", "license": "NVIDIA Open Model Agreement", "commercialUse": "Commercial deployment is governed by NVIDIA’s custom Open Model Agreement; users must review its conditions.", "openness": [ { "key": "weights", "status": "available", "detail": "Official model weights are downloadable from the linked model release.", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" }, { "key": "architecture", "status": "available", "detail": "The official model page documents parameter activation, modalities, and configuration.", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" }, { "key": "inference", "status": "available", "detail": "NVIDIA publishes reference deployment and inference guidance.", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" }, { "key": "trainingCode", "status": "partial", "detail": "Nemotron and NeMo tooling is public, but complete code for this exact training run was not established.", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" }, { "key": "recipe", "status": "partial", "detail": "NVIDIA documents the model’s construction and reasoning behavior without a complete end-to-end reproduction recipe.", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" }, { "key": "dataInformation", "status": "partial", "detail": "Training approach information is published, but the full multimodal mixture is not established here.", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" }, { "key": "evaluation", "status": "partial", "detail": "Vendor evaluation results and methods are published on the official model page.", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" }, { "key": "commercialUse", "status": "partial", "detail": "NVIDIA explicitly permits commercial use under its custom agreement. This is permission with conditions, not a prohibition and not an OSI certification.", "sourceUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" } ], "sources": [ { "id": "nemotron3-omni-report", "title": "Nemotron 3 Nano Omni technical report", "url": "https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Omni-report.pdf", "type": "report" }, { "id": "nemotron3-omni-page", "title": "Nemotron 3 Nano Omni 30B-A3B Reasoning", "url": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning", "type": "institution" }, { "id": "nemotron3-family", "title": "Nemotron AI models", "url": "https://developer.nvidia.com/topics/ai/nemotron", "type": "institution" } ], "strengths": [ "Video and audio understanding", "Reasoning", "Efficient active parameter count", "Long context" ], "limitations": [ "Custom license", "Complete multimodal training corpus is not released", "Vendor benchmarks need independent reproduction" ], "quantizations": [ "Check the official model page for supported NVIDIA deployment formats" ], "memory": "No single minimum is asserted; requirements depend on precision, media length, and runtime.", "discoveryIds": ["nemotron-3-super", "llama-nemotron-reasoning"], "independentEvidence": "No independent reproduction of the Nano Omni benchmark suite is attached; performance statements remain vendor-reported.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/nemotron", "recordUrl": "https://brightaifuture.com/open-models/nemotron", "licenseSummary": "Custom or artifact-specific terms: read them. Commercial deployment is governed by NVIDIA’s custom Open Model Agreement; users must review its conditions.", "licenseUrl": "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning", "download": { "gb": 22.409034192, "label": "About 22.41 GB of selected documented weights", "checkpoint": "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4", "format": "Safetensors, official NVFP4", "sourceUrl": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4/tree/16993199e436da4ba75ddc410855f87e0d996ee6", "checkedAt": "2026-10-11", "scope": "NVIDIA-published NVFP4 variant of Omni 30B-A3B Reasoning. Original unqualified repository redirects to BF16; selected variant has its own repository/revision. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 22409034192, "revision": "16993199e436da4ba75ddc410855f87e0d996ee6", "fileCount": 3, "metadataUrl": "https://huggingface.co/api/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4/revision/16993199e436da4ba75ddc410855f87e0d996ee6?blobs=true", "modelCardUrl": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4/blob/16993199e436da4ba75ddc410855f87e0d996ee6/README.md", "licenseUrl": "https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-agreement/", "artifactLicense": "nvidia-open-model-agreement", "publisher": "maker" } }, { "id": "smollm", "name": "SmolLM3", "organization": "Hugging Face", "version": "SmolLM3-3B", "releaseDate": "2025-07-08", "dateEvents": [ { "kind": "initial-release", "date": "2025-07-08", "precision": "day", "label": "SmolLM3 release", "sourceUrl": "https://huggingface.co/blog/smollm3" } ], "summary": "A compact multilingual reasoning model whose engineering blueprint, mixtures, training frameworks, and alignment process are published.", "parameters": "3B", "totalBillions": 3, "activeBillions": 3, "architecture": "Dense decoder transformer with grouped-query attention and alternating NoPE layers.", "modalities": ["text"], "context": "128K tokens", "license": "Apache-2.0", "commercialUse": "Permitted by Apache-2.0; constituent training datasets have their own terms.", "openness": [ { "key": "weights", "status": "available", "detail": "Base and instruct/reasoning weights are public.", "sourceUrl": "https://huggingface.co/blog/smollm3" }, { "key": "architecture", "status": "available", "detail": "Architecture and ablations are described in detail.", "sourceUrl": "https://huggingface.co/blog/smollm3" }, { "key": "inference", "status": "available", "detail": "Transformers-compatible inference and prompting instructions are public.", "sourceUrl": "https://huggingface.co/blog/smollm3" }, { "key": "trainingCode", "status": "available", "detail": "Nanotron, Datatrove, and Lighteval code and configurations are linked.", "sourceUrl": "https://huggingface.co/blog/smollm3" }, { "key": "recipe", "status": "available", "detail": "The three-stage data mixture and training blueprint are published.", "sourceUrl": "https://huggingface.co/blog/smollm3" }, { "key": "dataInformation", "status": "available", "detail": "Sources and mixture percentages are disclosed by training stage.", "sourceUrl": "https://huggingface.co/blog/smollm3" }, { "key": "trainingData", "status": "partial", "detail": "The named public datasets are accessible, but the release is not a single immutable corpus bundle.", "sourceUrl": "https://huggingface.co/blog/smollm3" }, { "key": "evaluation", "status": "available", "detail": "Evaluation details and Lighteval tooling are published.", "sourceUrl": "https://huggingface.co/blog/smollm3" }, { "key": "commercialUse", "status": "available", "detail": "The model is Apache-2.0 licensed.", "sourceUrl": "https://huggingface.co/blog/smollm3" } ], "sources": [ { "id": "smollm3-blog", "title": "SmolLM3: smol, multilingual, long-context reasoner", "url": "https://huggingface.co/blog/smollm3", "type": "institution" }, { "id": "smollm3-card", "title": "SmolLM3-3B model card", "url": "https://huggingface.co/HuggingFaceTB/SmolLM3-3B", "type": "registry" } ], "strengths": [ "Small footprint", "Transparent recipe", "Long context", "Dual reasoning mode" ], "limitations": [ "Text-only", "Six explicitly supported languages", "Smaller capacity than frontier-scale models" ], "quantizations": [ "Community GGUF, MLX, ONNX, and bitsandbytes variants exist" ], "memory": "The base parameter size makes consumer deployment practical when quantized; no single official minimum is asserted.", "discoveryIds": [], "independentEvidence": "No independent capability study is attached; the release’s reproducibility artifacts make such study practical.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/smollm", "recordUrl": "https://brightaifuture.com/open-models/smollm", "licenseSummary": "Commercial use permitted under Apache-2.0; review accompanying terms.", "licenseUrl": "https://huggingface.co/blog/smollm3", "download": { "gb": 1.92, "label": "1.92 GB listed GGUF file", "checkpoint": "SmolLM3-3B / SmolLM3-Q4_K_M.gguf", "format": "GGUF Q4_K_M", "sourceUrl": "https://huggingface.co/ggml-org/SmolLM3-3B-GGUF/tree/main", "checkedAt": "2026-10-11", "publisher": "conversion", "scope": "Selected ggml-org conversion of Hugging Face SmolLM3, not a maker-published artifact. This 1.92 GB listing is rounded, not total runtime memory or a census of all quantizations." } }, { "id": "molmo", "name": "Molmo 2", "organization": "Allen Institute for AI (Ai2)", "version": "Molmo2-4B, Molmo2-8B, and Molmo2-O-7B", "releaseDate": null, "dateEvents": [ { "kind": "initial-release", "date": "2025-12", "precision": "month", "label": "Molmo 2 release (month documented; day not established)", "sourceUrl": "https://allenai.org/media-center" } ], "releaseNotes": "Ai2’s primary media record establishes December 2025, not a release day; this field remains null rather than inventing one.", "summary": "A family of open vision-language models for images, multiple images, and video, with explicit spatial outputs such as pointing and object tracking.", "parameters": "4B, 7B, and 8B variants", "totalBillions": 8, "activeBillions": 8, "architecture": "Vision-language models with variants specialized for visual understanding and spatial grounding.", "modalities": [ "text", "image", "video", "pointing coordinates", "tracking coordinates" ], "context": "Varies by checkpoint and training stage; use the versioned model card.", "license": "Apache-2.0 repository and model terms; some source datasets require separate agreements", "commercialUse": "The released code and model artifacts use Apache-2.0, but dataset terms must be reviewed separately.", "openness": [ { "key": "weights", "status": "available", "detail": "Multiple official checkpoints are released after pre-training, supervised fine-tuning, and long-context fine-tuning.", "sourceUrl": "https://github.com/allenai/molmo2" }, { "key": "architecture", "status": "available", "detail": "The repository and technical report document model design and variant differences.", "sourceUrl": "https://github.com/allenai/molmo2" }, { "key": "inference", "status": "available", "detail": "Inference and demo code is included in the official repository.", "sourceUrl": "https://github.com/allenai/molmo2" }, { "key": "trainingCode", "status": "available", "detail": "Training scripts for the documented stages are public.", "sourceUrl": "https://github.com/allenai/molmo2" }, { "key": "recipe", "status": "available", "detail": "The repository documents staged training and links the technical report.", "sourceUrl": "https://github.com/allenai/molmo2" }, { "key": "dataInformation", "status": "available", "detail": "Training datasets and stage-specific inputs are identified.", "sourceUrl": "https://github.com/allenai/molmo2" }, { "key": "trainingData", "status": "partial", "detail": "Most dataset download scripts are supplied, while some datasets require separate agreements.", "sourceUrl": "https://github.com/allenai/molmo2" }, { "key": "evaluation", "status": "available", "detail": "Evaluation code and benchmark details are released with the repository.", "sourceUrl": "https://github.com/allenai/molmo2" }, { "key": "commercialUse", "status": "partial", "detail": "Apache-2.0 covers released artifacts, while source datasets retain separate terms.", "sourceUrl": "https://github.com/allenai/molmo2" } ], "sources": [ { "id": "molmo2-media-release", "title": "Ai2 media center — Molmo 2 release month", "url": "https://allenai.org/media-center", "type": "institution" }, { "id": "molmo2-repo", "title": "Molmo 2 repository", "url": "https://github.com/allenai/molmo2", "type": "repository" }, { "id": "molmo2-readme", "title": "Molmo 2 model and training documentation", "url": "https://github.com/allenai/molmo2/blob/main/README.md", "type": "report" } ], "strengths": [ "Video understanding", "Multi-image reasoning", "Pointing", "Object tracking", "Open training scripts" ], "limitations": [ "Some training datasets need separate agreements", "Spatial outputs still require task-specific evaluation before deployment" ], "quantizations": [ "No canonical quantized checkpoint is asserted in this record" ], "memory": "Requirements vary from the 4B to 8B variants and with video length; no single minimum is asserted.", "discoveryIds": [], "independentEvidence": "This record does not attach an independent reproduction of Molmo 2 results; it highlights that the released artifacts support such work.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/molmo", "recordUrl": "https://brightaifuture.com/open-models/molmo", "licenseSummary": "Commercial use permitted under Apache-2.0; review accompanying terms.", "licenseUrl": "https://github.com/allenai/molmo2", "download": { "gb": 19.403574432, "label": "About 19.40 GB of selected documented weights", "checkpoint": "allenai/Molmo2-4B", "format": "Safetensors, float32", "sourceUrl": "https://huggingface.co/allenai/Molmo2-4B/tree/042abfa7a38879a376cec03d949eff0aefaa0600", "checkedAt": "2026-10-11", "scope": "Molmo2-4B only, not 8B or O-7B. Four float32 shards include vision components, so 4B label is not whole tensor parameter count. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 19403574432, "revision": "042abfa7a38879a376cec03d949eff0aefaa0600", "fileCount": 4, "metadataUrl": "https://huggingface.co/api/models/allenai/Molmo2-4B/revision/042abfa7a38879a376cec03d949eff0aefaa0600?blobs=true", "modelCardUrl": "https://huggingface.co/allenai/Molmo2-4B/blob/042abfa7a38879a376cec03d949eff0aefaa0600/README.md", "licenseUrl": "https://huggingface.co/allenai/Molmo2-4B/blob/042abfa7a38879a376cec03d949eff0aefaa0600/README.md", "artifactLicense": "apache-2.0", "publisher": "maker" } }, { "id": "kimi", "name": "Kimi K2.5", "organization": "Moonshot AI", "version": "Kimi-K2.5", "releaseDate": "2026-01-27", "dateEvents": [ { "kind": "initial-release", "date": "2026-01-27", "precision": "day", "label": "Kimi K2.5 release", "sourceUrl": "https://github.com/MoonshotAI/kimi-help-center/blob/master/en-US/agent/swarm.md" }, { "kind": "update", "date": "2026-04-20", "precision": "day", "label": "Kimi K2.6 release noted by Moonshot AI", "sourceUrl": "https://github.com/MoonshotAI/kimi-help-center/blob/master/en-US/agent/swarm.md" } ], "releaseNotes": "K2.5 remains the reference checkpoint here; Moonshot AI’s cited documentation notes a later K2.6 release.", "summary": "A trillion-parameter visual agentic model with a comparatively small active footprint and public weights and code.", "parameters": "1T total / 32B active, plus a 400M vision encoder", "totalBillions": 1000, "activeBillions": 32, "architecture": "Multimodal mixture-of-experts transformer with MoonViT vision encoding.", "modalities": ["text", "image", "video"], "context": "256K tokens", "license": "Modified MIT License", "commercialUse": "Generally permitted, with an additional display requirement for products above the license’s stated revenue or user thresholds.", "openness": [ { "key": "weights", "status": "available", "detail": "Official model weights are downloadable.", "sourceUrl": "https://huggingface.co/moonshotai/Kimi-K2.5" }, { "key": "architecture", "status": "available", "detail": "Configuration and architecture details are public.", "sourceUrl": "https://huggingface.co/moonshotai/Kimi-K2.5" }, { "key": "inference", "status": "available", "detail": "The official repository provides code and deployment examples.", "sourceUrl": "https://github.com/MoonshotAI/Kimi-K2.5" }, { "key": "trainingCode", "status": "not-established", "detail": "Complete training code was not confirmed.", "sourceUrl": "https://huggingface.co/moonshotai/Kimi-K2.5" }, { "key": "recipe", "status": "partial", "detail": "The paper describes training and agentic methods without a complete reproducible recipe.", "sourceUrl": "https://huggingface.co/moonshotai/Kimi-K2.5" }, { "key": "dataInformation", "status": "partial", "detail": "Training strategy is described without release of the full mixture.", "sourceUrl": "https://huggingface.co/moonshotai/Kimi-K2.5" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://huggingface.co/moonshotai/Kimi-K2.5" }, { "key": "evaluation", "status": "partial", "detail": "Vendor evaluation protocols and results are published.", "sourceUrl": "https://huggingface.co/moonshotai/Kimi-K2.5" }, { "key": "commercialUse", "status": "restricted", "detail": "The modified MIT license adds a display requirement at defined scale thresholds.", "sourceUrl": "https://huggingface.co/moonshotai/Kimi-K2.5" } ], "sources": [ { "id": "kimi25-release-note", "title": "Moonshot AI Kimi K2.5 release note", "url": "https://github.com/MoonshotAI/kimi-help-center/blob/master/en-US/agent/swarm.md", "type": "repository" }, { "id": "kimi25-repo", "title": "Kimi K2.5 official repository", "url": "https://github.com/MoonshotAI/Kimi-K2.5", "type": "repository" }, { "id": "kimi25-card", "title": "Kimi K2.5 model card", "url": "https://huggingface.co/moonshotai/Kimi-K2.5", "type": "registry" }, { "id": "kimi25-paper", "title": "Kimi K2.5: Visual Agentic Intelligence", "url": "https://arxiv.org/abs/2602.02276", "type": "paper" } ], "strengths": [ "Visual agent workflows", "Coding", "Long context", "Sparse activation" ], "limitations": [ "Extremely large total checkpoint", "Custom license condition at large commercial scale", "Training data is not released" ], "quantizations": [ "Official/community NVFP4, MXFP4, FP8, GGUF, and MLX variants are listed on the Hub" ], "memory": "The full checkpoint is server-class; quantized serving still requires careful multi-device planning.", "discoveryIds": [], "independentEvidence": "No independent reproduction is attached; benchmark statements remain attributed to Moonshot AI.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/kimi", "recordUrl": "https://brightaifuture.com/open-models/kimi", "licenseSummary": "Custom or artifact-specific terms: read them. Generally permitted, with an additional display requirement for products above the license’s stated revenue or user thresholds.", "licenseUrl": "https://huggingface.co/moonshotai/Kimi-K2.5", "download": { "gb": 595.177988208, "label": "About 595.18 GB of selected documented weights", "checkpoint": "moonshotai/Kimi-K2.5", "format": "Safetensors, native INT4", "sourceUrl": "https://huggingface.co/moonshotai/Kimi-K2.5/tree/4d01dfe0332d63057c186e0b262165819efb6611", "checkedAt": "2026-10-11", "scope": "All 64 native INT4 checkpoint shards. Modified MIT, not unqualified MIT. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 595177988208, "revision": "4d01dfe0332d63057c186e0b262165819efb6611", "fileCount": 64, "metadataUrl": "https://huggingface.co/api/models/moonshotai/Kimi-K2.5/revision/4d01dfe0332d63057c186e0b262165819efb6611?blobs=true", "modelCardUrl": "https://huggingface.co/moonshotai/Kimi-K2.5/blob/4d01dfe0332d63057c186e0b262165819efb6611/README.md", "licenseUrl": "https://huggingface.co/moonshotai/Kimi-K2.5/blob/4d01dfe0332d63057c186e0b262165819efb6611/LICENSE", "artifactLicense": "modified-mit", "publisher": "maker" } }, { "id": "glm", "name": "GLM-5.3 Flash", "organization": "Z.ai", "version": "GLM-5.3-Flash", "releaseDate": "2026-08-26", "dateEvents": [ { "kind": "initial-release", "date": "2026-08-26", "precision": "day", "label": "GLM-5.3-Flash release", "sourceUrl": "https://github.com/zai-org/GLM-V" } ], "summary": "A natively multimodal sparse GLM checkpoint released under MIT terms for long-context agent and coding work.", "parameters": "320B total / 18B active", "totalBillions": 320, "activeBillions": 18, "architecture": "Mixture-of-experts transformer with hybrid sparse and linear attention.", "modalities": ["text", "image"], "context": "1M tokens", "license": "MIT", "commercialUse": "Permitted by the MIT license for this Flash checkpoint; other GLM 5.3 variants have different terms.", "openness": [ { "key": "weights", "status": "available", "detail": "Official weights are downloadable.", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash" }, { "key": "architecture", "status": "available", "detail": "Configuration and attention design are documented.", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash" }, { "key": "inference", "status": "available", "detail": "Official examples and a vLLM recipe are public.", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash" }, { "key": "trainingCode", "status": "not-established", "detail": "Complete training code was not confirmed.", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash" }, { "key": "recipe", "status": "partial", "detail": "Core methods are described without a complete reproduction recipe.", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash" }, { "key": "dataInformation", "status": "not-established", "detail": "A sufficiently detailed data mixture was not confirmed.", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash" }, { "key": "evaluation", "status": "partial", "detail": "Vendor results and evaluation information are published.", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash" }, { "key": "commercialUse", "status": "available", "detail": "The Flash checkpoint is MIT-licensed.", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash" } ], "sources": [ { "id": "glm53-release-note", "title": "Z.ai GLM-V release note", "url": "https://github.com/zai-org/GLM-V", "type": "repository" }, { "id": "glm53-card", "title": "GLM-5.3-Flash model card", "url": "https://huggingface.co/zai-org/GLM-5.3-Flash", "type": "registry" }, { "id": "glm53-vllm", "title": "vLLM GLM-5.3-Flash recipe", "url": "https://github.com/vllm-project/recipes/blob/main/models/zai-org/GLM-5.3-Flash.yaml", "type": "repository" } ], "strengths": [ "Long context", "Native multimodality", "Sparse inference", "Permissive checkpoint license" ], "limitations": [ "Training data is not disclosed in enough detail", "Large server-class checkpoint" ], "quantizations": [ "Runtime-specific community quantizations exist; the canonical release should be checked first" ], "memory": "No reliable universal minimum is stated; the 320B checkpoint requires distributed or quantized serving.", "discoveryIds": [], "independentEvidence": "The vLLM recipe corroborates deployability, not vendor capability claims.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/glm", "recordUrl": "https://brightaifuture.com/open-models/glm", "licenseSummary": "Commercial use permitted under MIT; retain the required notice.", "licenseUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash", "download": { "gb": 328.337455672, "label": "About 328.34 GB of selected documented weights", "checkpoint": "zai-org/GLM-5.3-Flash", "format": "Safetensors, FP8", "sourceUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash/tree/eb9eb208eb0d988989d07a6a12d0fdeb5f52574a", "checkedAt": "2026-10-11", "scope": "All 62 checkpoint shards; FP8 with unconverted modules as described by configuration. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 328337455672, "revision": "eb9eb208eb0d988989d07a6a12d0fdeb5f52574a", "fileCount": 62, "metadataUrl": "https://huggingface.co/api/models/zai-org/GLM-5.3-Flash/revision/eb9eb208eb0d988989d07a6a12d0fdeb5f52574a?blobs=true", "modelCardUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash/blob/eb9eb208eb0d988989d07a6a12d0fdeb5f52574a/README.md", "licenseUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash/blob/eb9eb208eb0d988989d07a6a12d0fdeb5f52574a/LICENSE", "artifactLicense": "mit", "publisher": "maker" } }, { "id": "evo", "name": "Evo 2", "organization": "Arc Institute and collaborators", "version": "Evo 2 7B and 40B", "releaseDate": "2025-02-19", "dateEvents": [ { "kind": "initial-release", "date": "2025-02-19", "precision": "day", "label": "Evo 2 release", "sourceUrl": "https://arcinstitute.org/news/evo2" } ], "summary": "A biological foundation model trained across genomes for sequence understanding and design, released with code, weights, and OpenGenome2 data.", "parameters": "7B and 40B", "totalBillions": 40, "activeBillions": 40, "architecture": "StripedHyena 2 sequence model for million-token genomic contexts.", "modalities": ["DNA sequence"], "context": "Up to 1M nucleotides", "license": "Apache-2.0 for code and model artifacts; dataset terms must be checked separately", "commercialUse": "Model code and weights are permissively licensed; biological datasets and downstream applications require separate diligence.", "openness": [ { "key": "weights", "status": "available", "detail": "7B and 40B model parameters are released.", "sourceUrl": "https://arcinstitute.org/manuscripts/Evo2.pdf" }, { "key": "architecture", "status": "available", "detail": "The paper and code describe StripedHyena 2.", "sourceUrl": "https://arcinstitute.org/manuscripts/Evo2.pdf" }, { "key": "inference", "status": "available", "detail": "Official inference and scoring code is public.", "sourceUrl": "https://github.com/ArcInstitute/evo2" }, { "key": "trainingCode", "status": "available", "detail": "Pretraining, midtraining and fine-tuning code is released separately in Savanna, as linked by the Evo 2 paper.", "sourceUrl": "https://github.com/zymrael/savanna" }, { "key": "recipe", "status": "available", "detail": "The paper documents model and training methodology.", "sourceUrl": "https://arcinstitute.org/manuscripts/Evo2.pdf" }, { "key": "dataInformation", "status": "available", "detail": "OpenGenome2 construction is documented.", "sourceUrl": "https://arcinstitute.org/manuscripts/Evo2.pdf" }, { "key": "trainingData", "status": "available", "detail": "OpenGenome2 is released for research use under its stated terms.", "sourceUrl": "https://arcinstitute.org/manuscripts/Evo2.pdf" }, { "key": "evaluation", "status": "available", "detail": "The paper publishes biological evaluations and methods.", "sourceUrl": "https://arcinstitute.org/manuscripts/Evo2.pdf" }, { "key": "commercialUse", "status": "partial", "detail": "Code and weights are permissive, while dataset and application constraints are not reducible to one permission.", "sourceUrl": "https://arcinstitute.org/manuscripts/Evo2.pdf" } ], "sources": [ { "id": "evo2-announcement", "title": "Arc Institute Evo 2 announcement", "url": "https://arcinstitute.org/news/evo2", "type": "institution" }, { "id": "evo2-paper", "title": "Genome modeling and design across all domains of life with Evo 2", "url": "https://arcinstitute.org/manuscripts/Evo2.pdf", "type": "paper" }, { "id": "evo2-training", "title": "Savanna training code linked by Evo 2", "url": "https://github.com/zymrael/savanna", "type": "repository" }, { "id": "evo2-repo", "title": "Evo 2 repository", "url": "https://github.com/ArcInstitute/evo2", "type": "repository" } ], "strengths": [ "Genomic sequence modeling", "Variant scoring", "Long biological context", "Open data and code" ], "limitations": [ "Research model, not a clinical diagnostic", "Generated biological sequences require domain safety review" ], "quantizations": ["No canonical quantized release is asserted here"], "memory": "The 40B model is server-class; the 7B model broadens access but still depends on context length and precision.", "discoveryIds": [], "independentEvidence": "The peer-reviewed manuscript includes collaborator evaluations; no claim of clinical validation is made.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/evo", "recordUrl": "https://brightaifuture.com/open-models/evo", "licenseSummary": "Commercial use permitted under Apache-2.0; review accompanying terms.", "licenseUrl": "https://arcinstitute.org/manuscripts/Evo2.pdf", "download": { "gb": 13.7666212, "label": "About 13.77 GB of selected documented weights", "checkpoint": "arcinstitute/evo2_7b", "format": "PyTorch .pt checkpoint", "sourceUrl": "https://huggingface.co/arcinstitute/evo2_7b/tree/bda0089f92582d5baabf0f22d9fc85f3588f6b58", "checkedAt": "2026-10-11", "scope": "evo2_7b.pt long-context 7B only, not separate base or 40B; not loaded or executed. Specialist sequence model, not ordinary laptop chat recommendation. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 13766621200, "revision": "bda0089f92582d5baabf0f22d9fc85f3588f6b58", "fileCount": 1, "metadataUrl": "https://huggingface.co/api/models/arcinstitute/evo2_7b/revision/bda0089f92582d5baabf0f22d9fc85f3588f6b58?blobs=true", "modelCardUrl": "https://huggingface.co/arcinstitute/evo2_7b/blob/bda0089f92582d5baabf0f22d9fc85f3588f6b58/README.md", "licenseUrl": "https://huggingface.co/arcinstitute/evo2_7b/blob/bda0089f92582d5baabf0f22d9fc85f3588f6b58/README.md", "artifactLicense": "apache-2.0", "publisher": "maker" } }, { "id": "groot", "name": "Isaac GR00T", "organization": "NVIDIA", "version": "GR00T N1.7", "releaseDate": "2026-04-18", "dateEvents": [ { "kind": "early-access", "date": "2026-04-18", "precision": "day", "label": "GR00T N1.7 early-access release", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T/releases" } ], "releaseNotes": "The cited official release is labelled early access; this record does not imply a later general-availability date.", "summary": "An open vision-language-action model for generalized humanoid skills, accompanied by reference code, data tooling, and a robotics development stack.", "parameters": "Not stated in the reviewed N1.7 overview", "totalBillions": null, "activeBillions": null, "architecture": "Vision-language-action foundation model combining perception, language instruction, and robot action generation.", "modalities": ["text", "image", "video", "robot state", "robot action"], "context": "Not established as a language-token context window in the reviewed release.", "license": "NVIDIA Open Model License / repository-specific terms", "commercialUse": "Review the model and data licenses for the intended robot, dataset, and deployment.", "openness": [ { "key": "weights", "status": "available", "detail": "Pre-trained N1.7 weights are linked from the official repository.", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T" }, { "key": "architecture", "status": "available", "detail": "The VLA architecture and interfaces are documented.", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T" }, { "key": "inference", "status": "available", "detail": "Reference inference and robot integration code is public.", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T" }, { "key": "trainingCode", "status": "available", "detail": "Fine-tuning and training utilities are public.", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T" }, { "key": "recipe", "status": "partial", "detail": "Training and fine-tuning procedures are documented, but full-scale reproduction depends on diverse robot data.", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T" }, { "key": "dataInformation", "status": "available", "detail": "Embodiments and dataset sources are documented.", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T" }, { "key": "trainingData", "status": "partial", "detail": "Substantial open robotics datasets are provided, but not every source used is represented as one corpus.", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T" }, { "key": "evaluation", "status": "partial", "detail": "Technical evaluations are published; real-world transfer remains platform-dependent.", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T" }, { "key": "commercialUse", "status": "partial", "detail": "NVIDIA permits commercial use under its model terms; dataset-specific permissions remain separate. Custom terms do not by themselves prohibit commercial use.", "sourceUrl": "https://github.com/NVIDIA/Isaac-GR00T" } ], "sources": [ { "id": "groot-n17-release", "title": "Isaac GR00T N1.7 release", "url": "https://github.com/NVIDIA/Isaac-GR00T/releases", "type": "repository" }, { "id": "groot-repo", "title": "Isaac GR00T repository", "url": "https://github.com/NVIDIA/Isaac-GR00T", "type": "repository" }, { "id": "groot-platform", "title": "Isaac GR00T platform", "url": "https://developer.nvidia.com/isaac/gr00t", "type": "institution" } ], "strengths": [ "Humanoid manipulation", "Cross-embodiment fine-tuning", "Open robotics tooling" ], "limitations": [ "Robot results depend on hardware, calibration, and safety controls", "Real-world operation needs physical safeguards" ], "quantizations": [ "Deployment formats are hardware- and runtime-specific" ], "memory": "Hardware requirements vary by robot platform and inference stack; no universal minimum is asserted.", "discoveryIds": [], "independentEvidence": "The original GR00T N1 paper provides research evaluation; N1.7 deployment claims remain primarily NVIDIA-reported.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/groot", "recordUrl": "https://brightaifuture.com/open-models/groot", "licenseSummary": "Custom or artifact-specific terms: read them. Review the model and data licenses for the intended robot, dataset, and deployment.", "licenseUrl": "https://github.com/NVIDIA/Isaac-GR00T", "download": { "gb": 6.910499416, "label": "About 6.91 GB of selected documented weights", "checkpoint": "nvidia/GR00T-N1.7-3B", "format": "Safetensors, BF16", "sourceUrl": "https://huggingface.co/nvidia/GR00T-N1.7-3B/tree/2fc962b973bccdd5d8ce4f67cc63b264d6886495", "checkedAt": "2026-10-11", "scope": "General checkpoint only, excludes task fine-tunes, scheduler.pt and training_args.bin. Preserve both card and bundled license references. Robotics specialist, not ordinary laptop chat recommendation. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 6910499416, "revision": "2fc962b973bccdd5d8ce4f67cc63b264d6886495", "fileCount": 2, "metadataUrl": "https://huggingface.co/api/models/nvidia/GR00T-N1.7-3B/revision/2fc962b973bccdd5d8ce4f67cc63b264d6886495?blobs=true", "modelCardUrl": "https://huggingface.co/nvidia/GR00T-N1.7-3B/blob/2fc962b973bccdd5d8ce4f67cc63b264d6886495/README.md", "licenseUrl": "https://huggingface.co/nvidia/GR00T-N1.7-3B/blob/2fc962b973bccdd5d8ce4f67cc63b264d6886495/LICENSE", "artifactLicense": "NVIDIA Open Model License Agreement (card); NVIDIA License (bundled file)", "publisher": "maker" } }, { "id": "voxtral", "name": "Voxtral Realtime", "organization": "Mistral AI", "version": "Voxtral-Mini-4B-Realtime-2602", "releaseDate": "2026-02-04", "dateEvents": [ { "kind": "initial-release", "date": "2026-02-04", "precision": "day", "label": "Voxtral Realtime release", "sourceUrl": "https://mistral.ai/news/voxtral-transcribe-2/" } ], "summary": "A compact streaming speech-recognition model designed for low-latency transcription through a realtime API.", "parameters": "4B", "totalBillions": 4, "activeBillions": 4, "architecture": "Streaming audio encoder-decoder model integrated with Mistral inference runtimes.", "modalities": ["audio", "text"], "context": "Streaming sessions; a comparable language-token window is not asserted here.", "license": "Apache-2.0", "commercialUse": "Permitted by Apache-2.0.", "openness": [ { "key": "weights", "status": "available", "detail": "Official weights are downloadable.", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" }, { "key": "architecture", "status": "available", "detail": "Configuration and model card are public.", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" }, { "key": "inference", "status": "available", "detail": "The release documents vLLM’s Realtime API integration.", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" }, { "key": "trainingCode", "status": "not-established", "detail": "Complete training code was not confirmed.", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" }, { "key": "recipe", "status": "not-established", "detail": "A full reproducible training recipe was not confirmed.", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" }, { "key": "dataInformation", "status": "partial", "detail": "Capabilities and languages are described without a complete corpus inventory.", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" }, { "key": "trainingData", "status": "not-established", "detail": "No sufficiently specific public artifact was confirmed in this review.", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" }, { "key": "evaluation", "status": "partial", "detail": "Vendor speech evaluations are published in the model materials.", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" }, { "key": "commercialUse", "status": "available", "detail": "The official checkpoint is Apache-2.0 licensed.", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" } ], "sources": [ { "id": "voxtral-rt-release", "title": "Voxtral Transcribe 2 and Realtime release", "url": "https://mistral.ai/news/voxtral-transcribe-2/", "type": "institution" }, { "id": "voxtral-rt-card", "title": "Voxtral Mini 4B Realtime model card", "url": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602", "type": "registry" }, { "id": "mistral-changelog", "title": "Mistral model changelog", "url": "https://docs.mistral.ai/resources/changelogs", "type": "institution" } ], "strengths": [ "Streaming transcription", "Low-latency audio processing", "Compact size" ], "limitations": [ "Speech-focused rather than a general multimodal assistant", "Training data is not fully disclosed" ], "quantizations": [ "Runtime-supported formats should be verified against the official card" ], "memory": "The 4B size supports comparatively modest deployment, but streaming overhead and audio session length affect requirements.", "discoveryIds": [], "independentEvidence": "vLLM integration corroborates production serving support; transcription quality claims are not independently reproduced here.", "checkedAt": "2026-09-19", "recheckDue": false, "recordPath": "/open-models/voxtral", "recordUrl": "https://brightaifuture.com/open-models/voxtral", "licenseSummary": "Commercial use permitted under Apache-2.0; review accompanying terms.", "licenseUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602", "download": { "gb": 8.859462744, "label": "About 8.86 GB of selected documented weights", "checkpoint": "mistralai/Voxtral-Mini-4B-Realtime-2602", "format": "Safetensors, Mistral consolidated representation", "sourceUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602/tree/2769294da9567371363522aac9bbcfdd19447add", "checkedAt": "2026-10-11", "scope": "Single consolidated.safetensors. Excludes alternative model.safetensors (8,859,446,848 bytes); alternate representations are not complementary shards. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.", "bytes": 8859462744, "revision": "2769294da9567371363522aac9bbcfdd19447add", "fileCount": 1, "metadataUrl": "https://huggingface.co/api/models/mistralai/Voxtral-Mini-4B-Realtime-2602/revision/2769294da9567371363522aac9bbcfdd19447add?blobs=true", "modelCardUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602/blob/2769294da9567371363522aac9bbcfdd19447add/README.md", "licenseUrl": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602/blob/2769294da9567371363522aac9bbcfdd19447add/README.md", "artifactLicense": "apache-2.0", "publisher": "maker" } } ] }