Mistral releases Mistral Large 4's open weights: a trillion-parameter AI anyone can download

Picture a hospital’s IT team, or a small bank’s security desk. They want strong AI to check code for weak points, but patient or customer data cannot leave their building. Until now, the biggest models mostly lived on someone else’s servers.

On 29 October, French company Mistral AI released the weights of Mistral Large 4, its flagship model, nicknamed “Le Chonk”. Weights are the learned numbers that make a model work; with them, anyone can download the model and run it on their own computers. The model has about 1 trillion parameters, and it comes with a licence that is free for most uses.

Mistral released this short film when the model first appeared in preview:

Inside Mistral Large 4: one trillion parameters, three ways to think

Mistral first showed the model as a public preview earlier in October; the version released now had about three more weeks of training. Here is what Mistral and its Hugging Face model card describe:

• Size: about 1 trillion parameters in total. Only about 52 billion of them work on each word, which keeps it cheaper to run.
• What it handles: text and images in one model, for instructions, reasoning, coding and multi-step “agent” tasks. It reads up to 1 million tokens at once (about the length of several long books) and, Mistral says, more than 160 languages.
• Three thinking modes: Low for simple jobs, Medium for everyday work, and High, the default, for the hardest tasks.
• Where to get it: download from Hugging Face, or use it through Mistral Studio, the Mistral API and OpenRouter.
• Price through the API: US$0.70 per million input tokens and US$2.10 per million output tokens (checked on 30 October 2026, on Mistral’s site and OpenRouter).
• Licence: Mistral calls it the Mistral Ecosystem Open-Weight licence, a modified MIT licence. You can use, change, fine-tune and sell with it. The one condition: a business that sells access to the model as an online service needs a separate commercial licence from Mistral.

Mistral says it trained the model from scratch on 3,800 NVIDIA chips in its own data centres in Europe. It also says coding and security results improved since the preview; a full technical report is “coming soon”, so outside experts have not yet checked those claims.

The same tool that defends can also attack

Mistral is pitching the Mistral Large 4 cybersecurity model at defenders: finding, reproducing and fixing software flaws. Pierre Stock, Mistral’s vice-president of science, told TechCrunch that open weights give organisations access that “can’t be cut off by definition”.

The hard part is that open weights are open for everyone. Mistral says its built-in guardrails refuse 96.6% of malicious cyber requests on average, by its own measurement. Its own announcement also admits it sees “a potential for uplift to conduct real attacks”, and says it is changing its guardrails and will keep updating them.

This is the analysis worth sitting with: once weights are public, anyone with the skill can fine-tune them, and safety training does not always survive that. Another lab this month chose to trim its model’s hacking skills before an open release. Stock told TechCrunch the risks are real but “solvable”.

If you build software or protect a network, what changes

Le Chonk will not run on a home laptop; Mistral’s own setup example uses a server with eight high-end graphics chips. If you just want to try Mistral Large 4, the API, Mistral Studio or OpenRouter are the easier doors.

For companies, governments, banks and hospitals, the meaning of an open-weight AI model is control. Data can stay in your own country and your own servers, and no outside company can switch the model off. Developers can download Mistral Large 4 from Hugging Face and adapt it, as long as they don’t resell it as a hosted service.

Why openness counts for more than the size

On balance, this is a meaningful step. A model this large, with a licence this free, gives smaller countries and organisations a real choice about where their AI lives. That matters to anyone worried about depending on a few providers, a theme behind other recent open-weight releases.

The caution is just as real. The strongest claims are Mistral’s own, the technical report is not out, and the company itself says its guardrails are still a work in progress. The likelier outcome is that the model proves most useful to well-funded teams with the hardware and skills to run it safely.

Watch for two things next: the technical report, and whether independent security researchers confirm that Le Chonk helps defenders more than it helps attackers.