Mistral Hands Out a Trillion-Parameter AI Model You Can Run Yourself

Europe now has a frontier-scale AI model that businesses and governments can download, run on their own servers and keep entirely in house. On 29 October, Paris-based Mistral released the full weights of Mistral Large 4, a model with roughly one trillion parameters, and it quickly became the top story on Hacker News.

That matters for sovereignty. A company or ministry running the model on its own hardware does not have to send data to an outside provider, and no one else can switch off its access. Pierre Stock, Mistral's vice president of science, told TechCrunch that access "can't be cut off by definition because the model is open-weighted."

The model first appeared as a public preview earlier in October. Mistral says it spent the weeks since on more training aimed at coding and defensive cybersecurity, and the version released now is that improved model.

What you actually get

Mistral says the model reads text and images, works in more than 160 languages including every official EU language, and was trained from scratch in its own European data centers. Those are company claims. Running it is not a laptop job. A compressed version needs four top-end Nvidia B200 or B300 chips, according to The Next Web, and Mistral's own setup guides assume a server with eight GPUs.

Mistral also says the model ties the leading open model, China's GLM-5.3, on a coding benchmark while using 47% fewer tokens per task. It says the model also leads open models on cyber defense tests and refuses 96.6% of malicious cyber requests in its own tests. These benchmark figures come from Mistral and have not been independently verified. Mistral itself says it saw "a potential for uplift to conduct real attacks" in live use and is tightening its safeguards.

The model also offers three reasoning settings, low, medium and high, so users can trade depth of thinking for speed and cost.

The license catch

The license is a modified MIT license. Anyone can use the model, change it and sell products built on it. There is one condition. A business that offers the model itself as a hosted service, letting customers send it their own prompts or training data, must sign a separate commercial license with Mistral. Internal use is exempt.

If you don't want to run it yourself

Not every organization wants to buy servers. Mistral sells access through its own API at $0.70 per million input tokens and $2.10 per million output tokens, according to its announcement. Microsoft says it is the first cloud provider to offer the model, through Microsoft Foundry, with private deployment options. Builders who host it themselves can use common open tools such as vLLM and SGLang.

Mistral's announcement also quotes banks and security firms, including La Banque Postale, ABN AMRO and Booz Allen Hamilton, describing early work with the model. Those are partner endorsements, not independent reviews.

For European organizations weighing US cloud AI against control, the bar just dropped. The cost now is hardware, not permission. The next test is independent benchmarking and the full technical report Mistral says is coming soon.

Sources

1. Mistral Large 4 (Le Chonk) is now open weight (Mistral)
2. Mistral-Large-4-1T-A52B model card (Hugging Face)
3. Mistral Ecosystem Open-Weight License (Mistral)
4. Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity (TechCrunch)
5. Mistral releases open weights for its 1 trillion-parameter Large 4 model (The Next Web)
6. We're releasing the weights of Mistral Large 4 (Hacker News)

Photo: "Wikimedia Foundation Servers-8055 35" by Victor Grigas, via Wikimedia Commons, CC BY-SA 3.0. Graded and overlaid by PSI, shared under the same license.