Mistral has handed over the keys to its biggest model, and its pitch is that nobody can take them back. The French company has released the weights of Mistral Large 4 (ML4), a trillion-parameter model nicknamed "Le Chonk," so that businesses and governments can download it and run it themselves. The first customers it has in mind are security teams, who are rarely thrilled about sending sensitive systems and data to somebody else's servers.
The model was first introduced earlier this month. What arrived on Thursday is the part that lets you actually own it.
What was released
The weights are the numerical parameters that make the trained model work. With them in hand, an organization can deploy ML4 for its own purposes, on its own hardware. Mistral is distributing them on Hugging Face, and the model is also available through the company's API.
The coverage describes ML4 both as "open source" and as "open-weighted," which are not the same thing. What the report actually describes is the release of weights. It does not say whether training code or training data come with them.
The performance claims
Mistral's headline claim is that ML4 delivers "state-of-the-art performance" at a "fraction of the compute and capital committed by closed labs." The specifics are narrower and worth reading closely.
| Claim | Comparison | Detail |
|---|---|---|
| Comparable performance | GLM-5.3, a Chinese open-weight model previously considered among the best open-weight models in the world, particularly for cyber work | Mistral says ML4 matches it, not beats it |
| Lower token use | GLM-5.3 | 47 percent fewer tokens per task, which Mistral says makes ML4 significantly more economical |
| Agentic coding | All other open models | Mistral says ML4 ranks first on Artificial Analysis' Coding Agent Index |
All of these are the company's own statements. The report gives no independent test results and no pricing, so the economics claim rests on token counts alone. Fewer tokens per task is a pleasant thing to have on an invoice, but it is not the invoice.
Why cybersecurity
Open-weight developers like Mistral have long argued that businesses and governments should not depend on outside AI companies for access to advanced capabilities. That argument competes with closed-model providers such as OpenAI and Anthropic, and it gets sharper in security, where organizations want tight control over their data and systems.
Mistral's statement leans into it. ML4, the company said, "brings state-of-the-art sovereign cyber defense capabilities to enterprises and governments globally," and institutions in the enterprise and public sectors "can now download, deploy, and own frontier AI to detect, analyze, and respond to threats across code and IT infrastructure."
Pierre Stock, Mistral's Vice President of Science, made the practical case in an interview. A business or a government, he said, gets access to frontier intelligence to defend itself against cyberattacks, and it is "access that can't be cut off by definition because the model is open-weighted." In other words, a vendor cannot change its terms, its pricing or its mind after you have built a security workflow on top of it.
Stock also said the model could be used to "identify, coordinate, and respond in real time to major security incidents." Because the weights are open, he added, companies can run it on-premises inside their own infrastructure.
The attackers get it too
The awkward part of the story is that anyone can download open weights, including the people defenders are worried about. Stock addressed the dual-use problem from the other direction. AI is already helping security professionals find bugs at a much higher rate than before, he said, and he thinks it matters to give the same capability to everyone as soon as possible, to strengthen defenses. His reasoning was that some malicious actors might try to use their own technology to attack anyway.
It is a coherent argument, and it is the standard one for releasing powerful tools openly: the defenders are the ones who need a head start, and a head start only exists if you hand it out. The report does not record any safeguards Mistral attached to the release, so readers will have to decide for themselves how comfortable they are with the bet.
On rogue agents
Stock was also asked about recent incidents involving rogue AI agents. He acknowledged the risks but cautioned against overreacting. "We don't want to fearmonger people," he said. "We think it's a problem that is solvable. It's a hard problem to have good security."
The last sentence is the most candid one in the whole announcement.
The larger argument
Mistral's CEO, Arthur Mensch, has framed the release in political terms. AI, he said, should "be a tool for empowerment, not dominance." With ML4, the company says it is doubling down on that position.
The test will come from the people who download it. A trillion-parameter model that matches the best open competitor on cyber work while using 47 percent fewer tokens would be a real argument for open weights, if the numbers hold up outside Mistral's own announcement. Until then, the main thing that has changed is that the weights are now yours to run, and the nice thing about a model you own is that nobody can send you a notice about it.