Mistral gives its biggest model away — and warns about it itself
On Thursday 29 October the French AI company Mistral put the files for its biggest model up for free download. The model is called Mistral Large 4 — known internally as "Le Chonk" — and it is a language model, the kind that writes and understands text, with particular strength in two things: writing program code and working on IT security.
The model has been running as a preview since 6 October. What is new is that the files themselves can now be downloaded and the model run on your own equipment, instead of sending your data to Mistral's servers. The files sit on Hugging Face, the service where AI models are normally posted for download. The model has a trillion parameters — the numerical values a model is built from — and, on Mistral's own account, uses only 52 billion of them at a time.
Open weights are not open source
Here is the first contradiction. Mistral consistently writes "open weight". That means the files are freely available, and it is not the same as open source, where anyone may also resell the product. The licence, which the company calls the Mistral Ecosystem Open-Weight license and itself describes as a modified MIT licence, grants a free right to use, modify, fine-tune and make money from the model — with one exception: anyone who wants to sell access to the model as a service must first buy a separate agreement from Mistral. Which is precisely the business the large American providers live on.
The company's senior figures use a different word. Chief executive and co-founder Arthur Mensch tells Euronews that "open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors", and TechCrunch, in its own text, calls the model open source. The difference is not pedantry. It decides who is allowed to make money from the model.
It says yes where others say no
The sales pitch is security. Mistral stresses that the model handles defensive security tasks "without unnecessary refusals" — a jab at the American models, which turn down that kind of work because it can also be used to attack. A specialised version for penetration testing is in preview together with Orange Cyberdefense.
The admission is in the same announcement. Mistral writes that it sees "a potential for uplift to conduct real attacks", and that the company is in the process of changing the model's guardrails. Pierre Stock, vice president of science, tells TechCrunch that malicious actors may try to use the technology to attack — but also that access to an open model cannot be shut off from the outside.
All the security figures are Mistral's own and have not been checked independently: 51 percent on the measure known as the Artificial Analysis Cyber Index, and 96.6 percent of malicious requests refused. On a test of practical programming work, Terminal-Bench 4.0, the model went from 28 to 41 percent since the preview. Whether it leads on coding is something the sources disagree about: Mistral itself writes that it is "effectively tied" with the Chinese model GLM-5.3, while TechCrunch writes that it comes out ahead of all open models. According to Euronews the model was trained on 3,800 Nvidia cards in Mistral's European data centres.
Two days earlier Mistral opened a new connection to its models — built to OpenAI's recipe. On the company's shareholders, see who owns Europe's AI companies.
Sources
- Mistral — the company's own announcement of the release: the licence terms, the security and coding figures, the warning about misuse and the Orange Cyberdefense arrangement
- Mistral — the 6 October preview announcement, with the parameter count and the 52 billion active parameters
- TechCrunch — the date, the quotes from Arthur Mensch and Pierre Stock, and the claim that it leads among open models
- Euronews — Mensch on open source, the three reasoning settings, the Terminal-Bench figures and the detail about the 3,800 Nvidia cards