You can try Mistral Large 4 from 6 October, but you cannot download it. The preview runs through the API in Mistral Studio, and the weights, the company says, arrive by the end of the month.
Fact
The facts: on 6 October 2026 Mistral released a public preview of Mistral Large 4, unofficially ML4, very officially "le Chonk". The model is natively multimodal, with 1 trillion parameters, 49 billion of them active, and in the company's words it is its largest to date. The model card in Mistral's documentation gives slightly different numbers: 1.05 trillion parameters, 52 billion active and a 1.6 billion vision encoder. Mistral promises the weights by the end of the month, along with details on the architecture and more benchmarks; that is a plan, not a finished fact. According to Mistral the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centers in Europe, the preview is served on the same infrastructure, and a significant share of the data covers more than 160 languages, including every official EU language. The remaining claims are also Mistral's: that it is competitive with the strongest open models in the world and significantly ahead of any open-weight model developed in the US or Europe, and that in cybersecurity, finance and law it is state of the art among open models. On the Artificial Analysis Cyber Index, by its account, it ranks in the top five models, and on the test of reproducing and patching a real vulnerability it scores 82%, the highest of any model. Until the weights are released, the company is testing the model with leading cybersecurity specialists, vetted partners and state authorities, who use the same model with reduced moderation and expanded cyber capabilities.
The hardware is the most concrete part. Three thousand eight hundred GPUs in Mistral's own data centers in Europe, and a European deployment the company says it will run end to end, under European law. For a customer who has to say where the computation goes, that is the answer Mistral is offering.
Open is something else. The weights are not out yet, so for now you have an API, as with any closed provider. Under what license they will come, Mistral does not say yet: in the documentation the license field reads "Coming soon".
Until the weights are out, an open model is a promise.
One sentence gets lost among the benchmarks: until the weights are released, security specialists, vetted partners and state authorities use ML4 with less moderation. Who exactly, the page does not say. As for why the cyber capability comes with open weights, Mistral points to refusals: a provider's refusal can stop legitimate vulnerability research or incident response, and losing access mid-incident is itself a risk.
The tables are Mistral's. Try the API with a task from your own work, in security or in law. When the weights are out, the first thing we will open is the license.
Official primary sources
Mistral - Introducing Mistral Large 4, 06.10.2026Mistral Docs - Mistral Large 4 (model card), 06.10.2026The visual is generated code-art, no third-party images.