Mistral Large 4 open weights — free to run 1T parameters, but API resale needs a license

Mistral kept its end-of-October promise and released its 1-trillion-parameter model in full. French AI company Mistral posted the weights of Mistral Large 4 to Hugging Face on October 29, 2026. It first unveiled the model as an API preview on October 6 and said then that the weights would follow at the end of the month ('Mistral Large 4 unveiled'). Governments and companies can now download the model and run it inside their own data centers.

Mistral's stated rationale is security. Pierre Stock, Mistral's VP of science, told TechCrunch that because the weights are open, by definition no one can cut off access. That is the opposite of Google's choice the same week, when it opened its top security model, Gemini 4 Carbon, first to vetted defenders only ('What is trusted access').

1. What was released

The API price is far below the closed top-tier models we have covered. Claude Opus 5.5 costs $4 input and $20 output; Gemini 4 Carbon's launch price is $2 input and $10 output. Mistral claims it reached top-level performance with a fraction of the compute and capital that closed labs use.

2. The license — 'open', with one condition

Mistral attached a new set of terms, the Mistral Ecosystem Open-Weight License, which it describes as a modified MIT license.

This is not a fully permissive license like plain MIT or Apache 2.0. The key is the last row: a cloud company or AI hosting provider that wants to put the model up as-is and sell it by the token has to pay Mistral. For Mistral, it is a way to stop competitors from running an API business on its own model. It is a similar path to Meta's Muse Spark, which released weights but was not open source ('Meta Muse Spark goes open'). The difference between open weights and open source is explained in 'What is an open-source model'.

3. Performance — top of the open models, still behind closed ones

Mistral spent the three weeks after the preview refining the model, and the 13-point gain on Terminal-Bench is the result. Notably, the comparisons are mostly against Chinese open models (GLM-5.3, Kimi, DeepSeek). Set against the claim that Kimi K3.1 Pro, released the previous week, came within 0.7 points of Opus 5.5 ('Kimi K3.1 Pro unveiled'), Large 4's strengths lie less in absolute scores than in token efficiency and its European origin.

4. Should anyone be able to download a security model?

Mistral says Large 4 can find vulnerabilities, reproduce them and patch them without breaking existing functionality. At the same time it acknowledged that, from observing real traffic, the model refused black-box attacks on real targets but has the potential to uplift real-world offensive capability. It said it is therefore adjusting safeguards.

Stock said he did not want to spread fear and called it a solvable problem. His logic is that attackers will use similar technology anyway, so defenders should get it first. Even so, Mistral opened its security-specialized version only as a preview to design partners such as Orange Cyberdefense, under contract. It has drawn a line somewhere between full openness and trusted access.

5. Open questions / what is unconfirmed

• Active parameter count: The repository name says 52 billion (A52B); reports at the preview said about 49 billion. Mistral's official post gives no number.
• Hardware requirements: Mistral has not disclosed how many GPUs or how much memory it takes to run a 1-trillion-parameter model. Ways to shrink model size are covered in 'What is model quantization'.
• Safeguard adjustments: What additional requests will be blocked has not been disclosed. Because the weights are already out, any later adjustments apply only to Mistral's own API.
• Independent verification: All scores are Mistral's.

Mistral raised 3 billion euros in September ('Mistral raises 3 billion euros'). Along with the Canadian government becoming a shareholder in its domestic AI company Cohere the same week, the push for AI sovereignty outside the US continues in 'Canada takes C$1.6 billion stake in Cohere'.