Mistral Large 4 Is Europe's Answer to the Chinese Open-Weight Giants
A one-trillion-parameter model in public preview, with weights promised for 27 October and a careful pitch about cyber defence.
Mistral released Mistral Large 4 on 6 October as a public preview on its API, with open weights scheduled for 27 October. The model, which Mistral has nicknamed "le Chonk", has about one trillion total parameters and 49 billion active per token, reads text and images, and was trained on more than 160 languages including every official EU language. Mistral says it trained the model from scratch over roughly two months on about 4,000 Nvidia Grace Blackwell GPUs, The Next Web reported.
Mistral positions Large 4 as a third way between closed American models and open Chinese ones, according to TechCrunch. It is aimed at coding, agentic work and specialist domains such as cybersecurity, finance and chip design.
Where it actually lands
Independent scoring puts Large 4 at 38 on the Artificial Analysis Intelligence Index, level with GPT-6 Luna at max effort and a point behind DeepSeek V4.1 Flash. That makes it the strongest model built outside the US and China by that measure, which is a real achievement for a European lab. It is also an honest reminder that it is not a frontier model in the sense Opus 5.5 or GPT-6 Astra are.
Mistral's own preliminary numbers are more flattering, and selective. It reports 62 per cent on DeepSWE v1.1, one point ahead of Zhipu's GLM-5.3; 67 per cent on a finance benchmark, level with DeepSeek-V4-Pro; and 73 per cent on a visual grounding test against 68 per cent for GPT-6 Astra. Each comparison picks a rival it beats or matches on a narrow task. That is normal launch practice, but it is not evidence of overall parity.
The improvement over Large 3 is clearer in document work. The API now accepts 100 images per request, up from eight, and Artificial Analysis credits that with an 18-point gain on document reasoning. For anyone processing scanned contracts, invoices or forms, that change may matter more than any coding score.
Smaller, and arguably smarter about it
At around a trillion parameters, Large 4 is less than half the size of Kimi K3 or Qwen3.8-Max. TechCrunch reports it was trained on two to three times fewer GPUs than its Chinese competitors. If it holds its own at that scale, it will be cheaper to host once the weights are out, which matters for the European banks, governments and manufacturers Mistral is courting. A trillion parameters is still multi-node territory, but it is a smaller cluster than its rivals need.
Pricing is not dramatic. Artificial Analysis lists $1.36 per million input tokens and $4.18 per million output, with a 50 per cent launch discount for two weeks, and notes that cost per task remains above some open-weight competitors such as GLM-5.3-Flash.
The cyber question
The three-week gap between API and weights is deliberate. Mistral says it will use the time for safety testing and to work with trusted partners and governments so the open weights can help defenders rather than attackers. Co-founder Guillaume Lample argued the model's cyber defence capabilities will let organisations protect themselves against attackers who jailbreak closed models.
That is a coherent position, and a tricky one. Open weights cannot be recalled, and safeguards that partners agree in October do not bind anyone who downloads the model in November. It is worth comparing with Anthropic routing most cyber requests away from Opus 5.5 unless customers are verified. The closed labs are tightening access to cyber capability at the same moment Mistral plans to publish it. Both cannot be right about the risk.
What to watch
- Whether the 27 October date holds, and whether the downloadable weights match the API model or ship stripped down, as Qwen's Max weights did.
- The licence the weights ship under. Apache 2.0 would be a strong statement; a revenue-triggered licence would follow the pattern Alibaba set with Qwen's Max weights.
- Independent cyber evaluations before release, ideally published.
Sources