Mistral has opened a public preview of Mistral Large 4, nicknamed Le Chonk, a trillion-parameter multimodal model with a particular pitch for cybersecurity. The company plans to release its weights on 27 October, after more testing, giving customers a chance to try the preview before they can run and modify the model themselves.
Mistral AI Watch analysis
What happened
Mistral says ML4 was trained on 4,000 Nvidia Grace Blackwell GPUs over two months in its European data centres. It is initially available through a moderated API. Selected testing partners will get a less restricted version with expanded cybersecurity capabilities, while Mistral continues safety testing and reinforcement learning ahead of the planned weights release.
The launch lands in a wider push for open-weight models that businesses can customise and host, rather than relying entirely on closed services. Axios reports that Mistral is presenting user control and cyber defence as central reasons to choose the model. The company says ML4 performs strongly on cybersecurity tasks, but those capability claims are not yet a substitute for independent results.
Why it matters
Once model weights are public, users can inspect and adapt a system, and are less dependent on one provider keeping an API available. The same freedom also makes safeguards harder for the maker to enforce. That trade-off matters especially for a model promoted for cybersecurity, where defensive use and misuse can sit uncomfortably close together.
Mistral’s staged approach is meant to create time for testing before broad release. The practical test will be what those evaluations show, and whether organisations can make good use of the model without treating “open” as a security feature in itself.
Our read
The interesting move is not simply that Mistral has built a large model; it is that the company is making control and cyber defence part of the case for opening it up. That case deserves a hearing, but the weights release is a milestone, not a verdict. Buyers should wait for published evaluations and test the model against their own workloads before handing it sensitive security work.
What to watch
- Whether Mistral releases ML4’s weights on 27 October as planned.
- What its safety testing and later benchmark results actually show.
- How the model performs in independent cybersecurity evaluations.
- Whether Mistral delivers the promised self-deployment and European data-residency options.
Discussion spark: For a cybersecurity model, should broad access to modifiable weights count as a security advantage, or does it make safeguards too difficult to enforce?
Sources and evidence
- "I don't want an oligopoly": New open-weight AI models mount a comeback against China (6 October 2026, 13:05 UTC)
not affiliated with or endorsed by Mistral AI