A Big Jump, Still Behind

Mistral released a public preview of Mistral Large 4 on 6 October 2026, a natively multimodal open-weight model with about one trillion parameters, 52 billion of them active. The API is available now through Mistral Studio and the weights are due by the end of the month.

ModelArtificial Analysis Intelligence Index
Mistral Large 39
Mistral Medium 3.514
Mistral Large 438
Claude Opus 5.5 (Max)58

The Decoder, citing the independent index, reports that the model also edges past GLM-5.2 from Z.ai but still trails the leading closed models and several Chinese open-weight ones. Mistral's own claim is narrower: the best open-weight model developed in the US or Europe. Trending Topics, also citing the index, reports 38.4 for the preview: the strongest open-weight model outside China, but only eighth among open-weight models, behind seven Chinese ones. Because Mistral is still training the preview, the score could change before the weights ship at the end of October.

Winning Where Closed Models Refuse

Mistral spends most of its announcement on security. On the Artificial Analysis Cyber Index it says Large 4 ranks among the top five models worldwide, and on one test that asks a model to reproduce a real vulnerability in open-source software and then patch it, it scores 82 percent, the highest of any model.

The reason is unusual. Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse the task. The Decoder's reading is that the test measures provider policy as much as capability. For a defender, proving a flaw is real is the first step of fixing it, and a model that declines to do it is not a tool.

Until the weights ship, Mistral is red-teaming the model with security firms, vetted partners and government agencies, who get the same version with reduced moderation and expanded cyber capabilities.

What Europe Gets: Residency And Weights

Mistral says Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centres in Europe, and that the public preview runs on that infrastructure. A European deployment is operated end to end, independently of other digital service providers and under European law.

It also says a significant share of the training data was multilingual, covering more than 160 languages including every official language of the European Union, and calls Large 4 the first milestone funded by its 3 billion euro Series D round.

These are Mistral's own claims, and the independent index shows the capability price of choosing sovereignty today: twenty points on general intelligence against the best closed model. Politico reported on the day of the launch that Ecosia, the Berlin search engine used by government services, is dropping Mistral for open-weight models served through Melious, citing quality and overloaded servers.

What To Do With A Sovereign Model

Choose by task rather than by flag. If your work is defensive security, incident response or regulated data that must stay in Europe, test Large 4 now through the European deployment. If your work needs the strongest general reasoning, keep a closed model and measure the gap on your own prompts.

Treat the weights as the real decision point. Self-hosting a model of this size is a hardware and operations commitment, so price a hosted European option first and revisit self-hosting after the weights are public and independent testers have run them.

Ask your closed-model vendors in writing what they refuse in security work and whether a verified-defender tier exists. Mistral has made that refusal gap a selling point, and you can use it in negotiation.