Mistral released the downloadable open weights for Large 4, its 1.05-trillion-parameter flagship model nicknamed “Le Chonk,” on Thursday, October 29 — turning the October 6 API preview into a model anyone can self-host. Why it matters: it is now the highest-scoring open-weight model developed outside China in Artificial Analysis testing, and it got much cheaper to run than the preview.
October 29, 2026
What happened
Mistral published the Large 4 weights on Hugging Face under its Mistral Ecosystem Open-Weight License, a modified MIT license. The model is free to use, modify, fine-tune and commercialize — with one catch: businesses that run a model-as-a-service offering need a separate commercial license from Mistral first. Internal enterprise use does not trigger that requirement. The weights can be served with vLLM or SGLang, and the model is also live through Mistral Studio and its API, OpenRouter, Mistral’s Vibe Code assistant, Microsoft Foundry (including private Azure deployments) and Nvidia-hosted endpoints with an optimized NIM container.
The details
Large 4 is a natively multimodal mixture-of-experts model: 1.05 trillion total parameters with about 52 billion active per request, a 1-million-token context window, and new low, medium and high reasoning modes. Mistral trained it on 3,800 Nvidia Grace Blackwell GPUs. The weights arrived two days later than the planned October 27 date, after roughly three extra weeks of reinforcement learning and feedback from government and corporate testers and security partners.
The independent numbers are more about efficiency than raw capability. Artificial Analysis scores the release at 26 on its Intelligence Index v5 at the high reasoning setting — one point above the preview and enough for the top spot among open-weight models built outside China, though still six points behind the open-weight leaders, DeepSeek V4.1 Flash and GLM-5.3. The bigger change is cost: $2.06 per Intelligence Index task versus $5.54 for the preview, a 63% drop, helped by API prices roughly halved to $0.70 per million input tokens and $2.10 per million output. On the Coding Agent Index it ranks second among open-weight models at 48, using about 47% fewer tokens per task than GLM-5.3, and its measured hallucination rate fell from 41% to 28%.
Cybersecurity is the other headline. Reuters reports Mistral designed the release to help organizations find and fix vulnerabilities on their own infrastructure, and the company calls it its most capable openly available model for that work — a field it says Chinese open models have dominated.
Why it matters
For subscribers, the practical story is control plus cost. A near-frontier model you can download cannot be repriced, rate-limited or switched off by a vendor, and this release narrowed the efficiency gap that made self-hosting hard to justify. The honest caveats: Chinese open-weight rivals still score higher and run cheaper per task in independent testing, Mistral’s own benchmark claims (such as its FrontierCode ranking) do not all match Artificial Analysis’s separate measurements, and a trillion-parameter model still needs serious hardware to self-host. But as an “everything AI, tested” takeaway: the open-weight frontier outside China just moved, and it moved on price as much as on power.
Frequently asked questions
Is Mistral Large 4 free to use commercially?
Mostly yes under its modified MIT-style license. The exception is model-as-a-service businesses, which need a separate commercial license from Mistral before using it commercially.
How is this different from the October 6 announcement?
October 6 was an API-only public preview. The October 29 release adds downloadable weights, three more weeks of training, new reasoning modes, more deployment platforms and API prices roughly half the preview’s.
Is Large 4 the best open-weight model?
It is the highest-scoring open-weight model developed outside China in Artificial Analysis’s Intelligence Index v5, but DeepSeek V4.1 Flash and GLM-5.3 score higher overall and cost less per task in the same testing.
Sources: Reuters, VentureBeat, Artificial Analysis (via VentureBeat), Unite.AI