Mistral previews Large 4
On 6 October 2026 Mistral AI launched a public preview of Mistral Large 4 on its Studio API: a natively multimodal model with 1 trillion parameters, 52 billion of them active, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres. The company promised to release the weights by the end of the month; until then it is red-teaming the model with cybersecurity partners and state authorities who get a version with reduced moderation. The preview costs $1.36 per million input tokens and $4.18 per million output tokens.
Why it matters
It is the largest model Mistral has named and the first it ties to its EUR 3 billion round; it also states a strategic bet: that open weights and self-hosting matter most in cybersecurity, where, Mistral argues, refusals by closed models block legitimate vulnerability work. That argument rests on Mistral's own tests and on a comparison in which it says Claude Opus 5.5 and GPT-6 Astra score near zero on one test because they refuse. The weights, which would let outsiders check the claims, had not been released when the page was read.
What the page says. Mistral Large 4 ("ML4", "le Chonk") is a "public preview", usable on Mistral Studio, with weights "by the end of the month". 1 trillion parameters, 52 billion active, natively multimodal; trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centres in Europe; a significant share of the training data multilingual, "more than 160 languages", including every official language of the EU. The page says the model uses the same training, customisation and RL environment that Mistral offers its customers through Mistral Forge, and calls it the first milestone on a roadmap funded by its "EUR 3 billion Series D", which Mistral calls the largest equity round ever raised by a European technology company. Price: $1.36 and $4.18 per million input and output tokens. The page names no licence and no context length. Until the weights are out, the model is being red-teamed "with cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities". Mistral's own results (the page's figures; third parties named where it names them). Cybersecurity: on the Artificial Analysis Cyber Index it ranks "among the top five models globally" and leads open-weight models developed outside China; on one test that asks a model to reproduce and patch a real vulnerability it scores 82%, "the highest of any model"; it solves 93% of the 40 Cybench challenges. Mistral writes that several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, "score near zero on the same test because they refuse to perform the task". Coding: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4; in a blind human evaluation with Surge AI it ranked second of five (3.74 on a 1-5 scale), behind Claude Opus 5 (4.22). Agents: 59.9% on AutomationBench (657 business workflows); 1,393 Elo on AA-Briefcase. Safety: it resists 93.3% of attacks on Lakera's B3 benchmark; 1.691 of a maximum of 2 on KORA. Knowledge work: Mistral says third-party evaluators (vals.ai) found that the model exceeds GPT-6 Astra on legal and financial tasks. What the record does not claim. These are Mistral's results; the weights were not yet available, so nobody else could repeat them, and the charts in the page are images that were not read. The claim about refusals is Mistral's reading of one test. The page gives no training cost.