Mistral launches open-source Mistral Large 4, details AI roadmap

Mistral AI has put Mistral Large 4 into public preview on its cloud platform, and the company is already talking about what comes next. The model is its most capable large language model so far, and Mistral says the weights will be released later this month. That open-source move is doing a lot of work here: the model arrives in the cloud first, but the company is promising the parts people can actually build on soon after.

Large 4 uses a mixture-of-experts setup with 1 trillion parameters, but only 49 billion are active for any given prompt. That matters because it makes the model less wasteful than turning the whole thing on every time someone asks a question. Mistral also says it can handle questions in more than 160 languages, which is the kind of detail that sounds broad until you remember how many models quietly trip over anything outside English.

The benchmark story is more mixed. Mistral says Large 4 landed in the top five on the AA Cyber Index, a set of tests for finding and fixing software vulnerabilities, and it did especially well on open-source patching with an 82% score. The company also says it beat GPT-6 Astra by 1% on Dense200, a computer-vision benchmark for spotting objects in images. But on the coding benchmarks that draw the most attention in the industry, Large 4 still trails frontier models like Astra even as it beats several open-source peers, including Qwen3.8 Max and DeepSeek V4 Pro.

Under the hood, Mistral says it trained Large 4 on 3,800 Grace Blackwell chips, each combining two Nvidia Blackwell GPUs with a CPU. It did not say how long that run took. It did say the training setup was unusual: engineers built a software stack that can run tens of thousands of rollouts at once, using code sandboxes, web search and tests to score the results. Those rollouts produced 33 billion tokens per day, and Mistral says it used just under half of them to refine the model.

The company also didn’t stop at one model. It says the rollouts and training workflow ran asynchronously, so one side could keep going without waiting for the other, and that it never paused the training run after the current version was created. Mistral expects bigger versions in the coming months, then a family of models tuned for specific uses. That’s the real tell here: not just a bigger model, but a machine for making more models.