Atlas AI Resources

← Back to AI News

AI models

Mistral released the weights of its trillion-parameter AI model. Here is what that means.

Published November 2, 2026

On October 29, the French AI company Mistral published the weights of Mistral Large 4, its biggest model, which it nicknames “Le Chonk.” That means anyone with enough computing power can download it and run it on their own machines instead of renting it through Mistral. Mistral pitches it at coding and cyber defense. An independent tester ranks it the strongest downloadable model built outside China, but still behind the best Chinese open models and well behind the top closed ones.

What happened

An AI model is, at bottom, a huge list of numbers called weights (or parameters) that it learned during training. Companies like OpenAI and Anthropic keep their weights secret and let you use the model only through their apps or paid connections. An open-weight model is one whose numbers are posted for download. Mistral put Large 4 out in a test version earlier in October and has now released an improved version’s weights on Hugging Face, a popular site for sharing AI models.

According to Mistral’s model card, Large 4 has about 1 trillion parameters but uses only about 49 billion of them at a time. That is a design called mixture of experts: the model is split into 128 specialist sections, and for each small chunk of text it switches on just four of them plus one shared section. It can read text and images, and it can hold about a million tokens (a token is a small piece of text, roughly part of a word) in view at once. Mistral says it trained the model from scratch on 3,800 Nvidia chips in its own European data centers.

The license lets people use, modify and sell products built on the model for free, with one catch: a business that resells access to the model itself as a hosted service needs a separate paid license from Mistral. If you would rather not run it yourself, Mistral charges $0.70 per million input tokens and $2.10 per million output tokens.

Why it matters to a normal person

Most people will never download a trillion-parameter model; Mistral’s own setup guide assumes a server with eight high-end graphics chips. The bigger effect is on the banks, hospitals, governments and software makers whose tools you use. With open weights, they can keep sensitive data on their own computers, adjust the model for their work, and keep using it even if a vendor changes its terms. Pierre Stock, Mistral’s vice president of science, told TechCrunch that access to an open-weight model “can’t be cut off by definition.”

Mistral is leaning hard on cybersecurity. It says Large 4’s built-in guardrails, the safety rules trained into the model, refused 96.6% of malicious hacking requests in its tests. A specialized cyber version is not public; Mistral is offering it only to vetted partners such as Orange Cyberdefense.

What it does not mean

It does not mean Large 4 is the best model available. Artificial Analysis, a company that runs its own tests on AI models, gives it 26 on its overall index. The top downloadable models, DeepSeek V4.1 Flash and GLM-5.3, both from Chinese labs, score 32. Google’s closed Gemini 4 Carbon, which is not yet public, tops the same index at 59.

It does not mean the coding and security numbers are settled. Claims like “best open-weight model for coding” and the 96.6% refusal rate are Mistral’s own, and its full technical report is still “coming soon.” Mistral also says that, watching real use since the preview, it saw “a potential for uplift to conduct real attacks” and is changing its guardrails. That points to the trade-off with any open model: once people download a copy, the company cannot recall it or update the safety rules inside it.

And it does not mean “open source” in the strict sense. You get the finished weights, but not the training data, and commercial hosting still needs Mistral’s permission.

Related: Google’s new Gemini 4 Carbon model goes to cyber defenders first. Here is why.

Sources