French AI company Mistral on Thursday, Oct. 29, released Mistral Large 4, its largest AI model, as an open-weight model: the model's trained files are published, so anyone can download it and run it on their own computers. Organizations can run it on their own systems instead of depending on Mistral's online service, which Mistral pitches to banks, governments and security teams. Mistral had promised the files by the end of October when it opened a preview to developers on Oct. 6, and it says the version it released is better, after three more weeks of training aimed at writing code and defending computer systems.
Large 4 has about 1 trillion parameters, the numbers a model tunes as it learns. It is a mixture-of-experts model, a design that switches on only part of the model for each piece of text, so it runs faster than its full size suggests. About 50 billion of its parameters work on each piece of text, but the whole model still has to fit in the computer's memory. It is not a model for a laptop: Mistral's own setup example runs it on a server with eight AI chips. It reads images as well as text, and Mistral says it trained the model from scratch in its own data centers in Europe.
Anyone may use, change and sell products built on it under Mistral's own license, with one condition: a business that runs AI models as an online service for others needs a separate commercial license from Mistral. People who don't want to run it themselves can use it through Mistral's online service, its Vibe coding tool or Microsoft's Foundry cloud service, among others. Mistral's online service now charges $0.70 for every million tokens the model reads and $2.10 for every million it writes, about half the preview's $1.36 and $4.18. Tokens are the small chunks of text, often parts of words, that AI models read and write.
All of the test results here come from Mistral, which chose the tests and the rival models to show. Mistral says Large 4 is now the best free-to-download model for writing code. On a test from the coding company Cognition, which asks whether the person who runs a software project would accept the changes a model makes, Mistral says it scored highest of the downloadable models it compared, though its score was 19.7%, under one in five. On another outside coding ranking, Mistral says Large 4 only tied the leading downloadable model, GLM-5.3, while using about half as many tokens. It also says graders comparing answers side by side, without knowing which model wrote them, preferred its code to that of GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, three leading downloadable models, and ranked it behind only Anthropic's Claude Opus 5.5 among the models tested.
Mistral is pitching the model hardest at security teams. It says Large 4 scored 51% on an outside test of finding, reproducing and fixing weaknesses in software, and does that defensive work without needless refusals, though in Mistral's own chart two other downloadable models, DeepSeek V4.1 Flash and MiMo-V2.6-Pro, scored higher on that test. Mistral says the model also turns down 96.6% of malicious hacking requests in its tests. Mistral says it has watched how people used the model for security and hacking work on its own service since the preview. It says it saw the model refuse to carry out attacks on live systems it was asked to break into, but it still sees a risk that the model could help people carry out real attacks, so it is changing the model's safety limits and will keep updating them in the coming months. Mistral has not said whether those changes will reach copies people have already downloaded. Pierre Stock, a Mistral vice president, said that once the model is out, "it's impossible to contain access," according to a news report, which adds that this means no one can cut users off, but anyone can use it.
Mistral also says that on a combined ranking from Artificial Analysis, a company that tests AI models, which pools many tests, Large 4 scores highest among downloadable models from Western companies but below the leading downloadable models from China. Mistral is offering security partners such as Orange Cyberdefense early access to a separate version built for testing a company's defenses by attacking them, with access vetted and limited to systems the partner owns or is allowed to test, and has not said whether that version will be released to anyone else.