{"schema":"postcutoff/event@1","as_of":"2026-10-30T13:42:00+01:00","url":"https://postcutoff.com/e/2026-10-29-mistral-large-4-open-weights/","md":"https://postcutoff.com/e/2026-10-29-mistral-large-4-open-weights/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-10-29-mistral-large-4-open-weights","date":"2026-10-29","date_precision":"day","short_title":"Mistral releases the weights of an improved Mistral Large 4 (\"Le Chonk\"), a 1T-parameter MoE, under a modified MIT licence","deck":null,"takeaway":null,"category":"model-release","category_label":"Model releases","importance":4,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":[{"n":1,"title":"Mistral AI: Mistral Large 4 (Le Chonk) is now open weight","url":"https://mistral.ai/news/mistral-large-4-oss/","type":"official","group":"primary","domain":"mistral.ai"},{"n":2,"title":"Hugging Face blog: Welcome Le Chonk, Mistral Large 4 on Hugging Face (Oct 29)","url":"https://huggingface.co/blog/mistral-large4","type":"docs","group":"primary","domain":"huggingface.co"},{"n":3,"title":"Hugging Face: mistralai/Mistral-Large-4-1T-A52B (FP8 weights)","url":"https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B","type":"code","group":"primary","domain":"huggingface.co"},{"n":4,"title":"Hugging Face: mistralai/Mistral-Large-4-1T-A52B-NVFP4","url":"https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B-NVFP4","type":"code","group":"primary","domain":"huggingface.co"},{"n":5,"title":"Mistral AI on X: 'Today, we release the weights of a new and improved Le Chonk' (~310k views)","url":"https://x.com/MistralAI/status/2115811242556137820","type":"official","group":"primary","domain":"x.com"},{"n":6,"title":"Artificial Analysis: Mistral Large 4 model page","url":"https://artificialanalysis.ai/models/mistral-large-4","type":"docs","group":"primary","domain":"artificialanalysis.ai"},{"n":7,"title":"Mistral AI on X: Responses API in public preview; serving DeepSeek V4.1 Flash and GLM 5.3 Flash (Oct 28)","url":"https://x.com/MistralAI/status/2115443720371814478","type":"official","group":"primary","domain":"x.com"},{"n":8,"title":"TechCrunch: Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity","url":"https://techcrunch.com/2026/10/29/mistral-releases-weights-for-its-trillion-parameter-ai-model-targeting-cybersecurity/","type":"press","group":"press","domain":"techcrunch.com"},{"n":9,"title":"Reuters: Mistral AI publicly releases Large 4, targets cybersecurity users","url":"https://www.reuters.com/legal/litigation/mistral-ai-publicly-releases-large-4-targets-cybersecurity-users-2026-10-29/","type":"press","group":"press","domain":"reuters.com"},{"n":10,"title":"ZDNET: Mistral Large 4 weights release (Techmeme headline: rivals GLM-5.3 on the AA Coding Agent Index)","url":"https://www.zdnet.com/innovation/mistral-large-4-weights-release/","type":"press","group":"press","domain":"zdnet.com"},{"n":11,"title":"Hacker News: We're releasing the weights of Mistral Large 4 (816 pts)","url":"https://news.ycombinator.com/item?id=50289418","type":"discussion","group":"reactions","domain":"news.ycombinator.com"}],"official":7,"filed":"2026-10-29","updated":"2026-10-29","orgs":["Mistral AI"],"title":"Mistral releases the weights of an improved Mistral Large 4 (\"Le Chonk\"), a 1T-parameter MoE, under a modified MIT licence","summary":"On Oct 29, 2026 Mistral AI published the weights of a further-trained Mistral Large 4 (\"Le Chonk\"), the ~1T-parameter, 49B-active MoE it previewed on Oct 6, on Hugging Face (FP8 and NVFP4) under its new Mistral Ecosystem Open-Weight licence, a modified MIT licence that requires model-as-a-service businesses to sign a commercial licence. It halved the API price to $0.70/$2.10 per 1M tokens and says the model leads open models on agentic coding; it also said monitoring of the preview's traffic showed \"a potential for uplift to conduct real attacks\" and that it is changing its guardrails.","key_facts":["Released: weights on Hugging Face as mistralai/Mistral-Large-4-1T-A52B (FP8) and mistralai/Mistral-Large-4-1T-A52B-NVFP4; GGUF quants from ggml-org (llama.cpp) and Unsloth on day one (Hugging Face blog, Oct 29)","Architecture (HF blog): ~1T total, 49B active parameters (52B with embeddings and output layers, hence 'A52B'); 72 layers, 128 experts with 4 routed + 1 shared per token, Multi-head Latent Attention with a DeepSeek Sparse Attention indexer (up to 2,048 keys per query), 1.6B Pixtral vision encoder, 1M-token context","Licence: 'Mistral Ecosystem Open-Weight license, a modified MIT license'; free use, modification, fine-tuning and commercial use, except that a licensee operating a Model-as-a-Service business needs a separate commercial licence first. Hacker News commenters called it 'weights available' rather than open weight (not OSI-approved); Large 3 had been Apache 2.0","Changes since the Oct 6 preview (Mistral): more RL, with 'Gains continue to accelerate, with no signs of saturation'; three reasoning modes Low (replaces 'None'), Medium and High (default). The HF blog says reasoning cannot be disabled","Mistral-reported results: Terminal-Bench 4.0 41% (preview 28%, with 30% fewer output tokens); Artificial Analysis Coding Agent Index 48%, 'effectively tied' with GLM-5.3 at 47% fewer tokens per task; Cognition FrontierCode 2.0 19.7% (#1 among open models); AA Cyber Index 51%; BrowseComp 84.5%; FinWorkBench 73.3%; Lakera B3 attack resistance 97.1%; 96.6% average refusal of malicious cyber requests. All self-reported","Independent: Artificial Analysis lists Mistral Large 4 (High) at 26 on Intelligence Index v5 (open-weight median for its size 23), 161 output tokens/s, ~$2.06 per index task; the HF blog says it is the fifth-best open-weight model on the index as of Oct 29","Price: $0.70 input / $0.07 cached / $2.10 output per 1M tokens on the Mistral API (id mistral-large-4-0), about half the preview's standard $1.36/$4.18","Availability: Mistral Studio and API, Hugging Face, OpenRouter (mistralai/mistral-large-4-0), NVIDIA build.nvidia.com and a NIM container, Microsoft Foundry (first cloud), vLLM, SGLang, TensorRT-LLM, llama.cpp","Hardware (HF blog): NVFP4 in vLLM needs Blackwell GPUs (example on 8 GPUs); the 2-bit GGUF runs on one 512 GB Apple Silicon Ultra machine (~15 tok/s generation on an M3 Ultra at Q2_K); FP8 on 8x H200 gives ~65 tok/s for a single request","Cyber: partners named include Booz Allen, Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet, NVIDIA, Orange Cyberdefense, Rubrik and Zscaler; ENISA 'is experimenting with closed and open weight models in order to better understand their cyber capabilities'; a more specialised cyber variant stays in preview for vetted design partners such as Orange Cyberdefense, 'subject to rigorous vetting and contractual safeguards'","Misuse note (Mistral): 'In the time since the preview, we carefully monitored cybersecurity-related use in our live traffic … we see a potential for uplift to conduct real attacks. We are modifying our guardrails accordingly, and will continue updating them in the months to come.'","Pierre Stock, Mistral's VP of Science (TechCrunch): the model can help organisations 'identify, coordinate, and respond in real time to major security incidents'; on recent rogue-agent incidents, 'We don't want to fearmonger people'. Arthur Mensch: AI should 'be a tool for empowerment, not dominance'","Reach: Mistral's X post had ~310k views by 20:30 CET on Oct 29; the Hacker News thread had 816 points and 218 comments","Same week: on Oct 28 Mistral put a Responses API for agents into public preview and began serving third-party open models (DeepSeek V4.1 Flash in preview, GLM 5.3 Flash) on its EU/US inference cloud"],"key_numbers":[],"tags":["llm","open-weights","moe","multimodal","europe","cybersecurity","agents","coding"],"science":null,"body_md":"## What happened\n\nMistral AI published the weights of Mistral Large 4 (\"Le Chonk\") on October 29, 2026, three weeks after the model went into\npublic preview (see 2026-10-06-mistral-large-4). The released checkpoint is not the preview: Mistral kept training it with\nreinforcement learning and says coding and defensive-cyber scores rose, for example Terminal-Bench 4.0 from 28% to 41%, with\nfewer output tokens. Reasoning now has three settings (Low, Medium, High); it can no longer be switched off.\n\nThe weights come in FP8 and NVFP4 on Hugging Face, with llama.cpp and Unsloth quantisations on the same day. They use a new\n\"Mistral Ecosystem Open-Weight license\", described by Mistral as \"a modified MIT license\": anyone may use, modify and sell\nproducts built on it, but a company that runs a model-as-a-service business must first sign a commercial licence. Hacker News\ncommenters said that makes it \"weights available\" rather than open source. The API price fell to $0.70 per million input\ntokens and $2.10 per million output tokens, roughly half the preview's list price.\n\nMistral pitches the model for agentic coding and for cyber defence, naming about ten security vendors as partners and saying\nthe EU cybersecurity agency ENISA is testing it. It also says that while the preview was live it \"carefully monitored\ncybersecurity-related use\" and saw \"a potential for uplift to conduct real attacks\", and that it is \"modifying our guardrails\naccordingly\". A more specialised cyber variant stays with vetted design partners such as Orange Cyberdefense. Guardrails that\nMistral runs on its own API do not apply to copies of the weights that others run themselves.\n\nBenchmarks in Mistral's post are self-reported. Artificial Analysis's own page puts the model at 26 on its Intelligence Index v5,\nabove the open-weight median for its size; Hugging Face calls it the fifth-best open-weight model on that index.\n\n## Why it matters\n\nIt is the largest open-weight model from a European lab. One Hacker News commenter argued that it \"opens a lot of doors\" for\nregulated US businesses that cannot run Chinese models and have relied on smaller US open models such as gpt-oss or Gemma. It\nalso shows a lab publishing weights while saying it saw uplift for real attacks in its own traffic, and keeping its strongest\ncyber variant gated.","disputed":[],"related":[{"id":"2026-10-28-irregular-glm-5-3-cyscenariobench-open-weight-record","url":"https://postcutoff.com/e/2026-10-28-irregular-glm-5-3-cyscenariobench-open-weight-record/","date":"2026-10-28","date_precision":"day","short_title":"Irregular: open-weight GLM-5.3 completes 3 of 13 multi-stage cyberattack scenarios, refuses none, and breaks instructions to reach correct answers","deck":null,"takeaway":"Open weights cannot be recalled or wrapped in safeguards once released, so each step toward complete attack chains in a downloadable model matters more than the same step in an API model.","category":"policy-safety","category_label":"Policy & safety","importance":3,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":4,"official":2,"filed":"2026-10-29","updated":"2026-10-29","orgs":["Irregular","Zhipu AI"]},{"id":"2026-10-06-mistral-large-4","url":"https://postcutoff.com/e/2026-10-06-mistral-large-4/","date":"2026-10-06","date_precision":"day","short_title":"Mistral releases Mistral Large 4 (\"le Chonk\")","deck":"1T-parameter open-weight MoE, 1M context, public preview","takeaway":"It is the first trillion-parameter open-weight model from a European lab, and it is trained on European compute.","category":"model-release","category_label":"Model releases","importance":4,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":21,"official":8,"filed":"2026-10-06","updated":"2026-10-29","orgs":["Mistral AI"]}],"people":[{"id":"arthur-mensch","name":"Arthur Mensch","url":"https://postcutoff.com/person/arthur-mensch/"}],"posts":[{"id":"x-mistralai-2115811242556137820","title":"“Today, we release the weights of a new and improved Le Chonk, advancing the open frontier…”","url":"https://postcutoff.com/p/x-mistralai-2115811242556137820/"},{"id":"x-mistralai-2115443720371814478","title":"“Today, we’re introducing the Responses API in public preview, made for building agents on…”","url":"https://postcutoff.com/p/x-mistralai-2115443720371814478/"}],"videos":[],"models":[{"id":"mistral-large-4","name":"Mistral Large 4","url":"https://postcutoff.com/m/mistral-large-4/"}],"changes":[{"date":"2026-10-29","type":"filed","text":"Created (Mistral post, Hugging Face blog, TechCrunch, Artificial Analysis, HN thread)"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-10-29","run":null,"sources_read":null,"updated":"2026-10-29","human_review":null,"version":{"date":"2026-10-29"}},"gaps":[{"model_id":"gpt-6-astra","name":"GPT-6 Astra","cutoff":"2026-04","days_after":182,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":121,"in_training_data":false},{"model_id":"gemini-4-argon","name":"Gemini 4 Argon","cutoff":"2026-03","days_after":212,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":151,"in_training_data":false}],"short_url":"https://postcutoff.com/s/mistral-weights-improved-mistral"}