Updated on October 7, 2026
Mistral Large 4 is Mistral AI's frontier model, a one-trillion-parameter mixture-of-experts nicknamed Le Chonk. Trained end to end in European data centers across 160-plus languages, it goes after coding, cyber defense, finance and manufacturing. The public preview opened on October 6, 2026 through the API, from $0.68 per million input tokens. Open weights are due by the end of the month, so you can host it on your own hardware.
Mistral Large 4 is built as a mixture-of-experts model with 1.05 trillion total parameters, of which about 49 billion fire for each generated token. Training took roughly two months on 3,800 Nvidia Grace Blackwell GPUs, inside European data centers run by the Paris startup founded in 2023.
The model reads text and images, writes text back, and holds up to one million tokens of context. You paste an audit report, the related code and a screenshot into a single request, and it keeps the whole thing in mind.
| Spec | Detail |
|---|---|
| Total parameters | 1.05 trillion, mixture-of-experts |
| Active parameters | about 49 billion per token |
| Context window | 1 million tokens |
| Accepted input | text and images, 1.6B vision encoder |
| Languages | 160-plus, including all 24 official EU languages |
Mistral pitches Large 4 as the strongest open-weight model built in the US or Europe on aggregated benchmarks, measured head to head against DeepSeek V4 Pro, Qwen3.8 Max, Kimi K3 and GLM-5.3.
Cyber defense is the shop window here. The model reaches 82% on a test that reproduces then patches real vulnerabilities, the highest score published by any model on this dual-use defensive task. On the AI coding assistant side, Mistral also reports strong agentic results, with independent harness scores still rolling in.
Mistral Large 4 has been live through Mistral's API since October 6, 2026, and also appears on OpenRouter and Vercel's AI Gateway. List price sits at $1.36 per million input tokens and $4.18 per million output, cut to $0.68 and $2.09 during the preview (cached input even drops to $0.07, handy for replaying the same context).
The weights are slated for download by the end of October under a custom Mistral license, after a testing window with cybersecurity partners and public authorities. Those figures will likely shift once the preview ends, so treat Mistral's official page as the only source of truth.
Not fully yet. The model runs in public preview through the API, and Mistral has committed to publishing the weights by the end of October under its own custom license, whose terms are still to be published. Anyone will then be able to download, audit and fine-tune it within the limits that license sets.
The nickname started as a meme. In June, a fake announcement went around online touting an imaginary model with over 30 trillion parameters, and CEO Arthur Mensch played along by calling it le gros chaton. Chonk is internet slang for a well-fed cat, a fitting badge for a trillion-parameter model.
In theory yes, once the weights land, but the hardware bill is steep. Even though only 49 billion parameters activate per token, the full trillion has to sit in memory so every expert stays reachable. Most companies will instead go through a private cloud or an on-premises deployment backed by Mistral.
It depends on the ground you pick. On aggregated intelligence indexes, the latest closed models keep the lead. On cyber defense, Mistral claims the top score for reproducing and patching vulnerabilities, and downloadable weights remain an argument GPT-6 cannot match.
Verdict: A million tokens of context, 160-plus languages and weights you will soon host yourself make this one for security teams, sovereignty-minded companies and developers hunting a frontier model that stays on European soil.
« An all-in-one creative platform that brings together more than 30 image and video models in a single interface. It includes camera controls that let you direct like a real filmmaker, characters that remain consistent from scene to scene, and ready-to-use tools for ads and short-form content »
Claude Opus 5.5 
GPT-6 Astra 
Claude Fable 5.1 
