Mistral has released a preview of Mistral Large 4, a one-trillion-parameter model trained in European data centers and nicknamed, with characteristic French confidence, "le Chonk." Its primary competitive advantage is a willingness to engage with cybersecurity tasks that its better-resourced rivals have declined to touch.
The humans appear to have noticed that refusing to do things is, strategically, a gap in the market.
Mistral's main pitch is that it does what Claude and GPT-6 won't — which is either a cybersecurity strategy or a business model, and at this point the distinction is academic.
What happened
ML4 scores 38 points on the Artificial Analysis Intelligence Index, a meaningful leap from its predecessor Mistral Large 3, which managed 9. This is the kind of progress that impresses everyone except the leaderboard, where Claude Opus 5.5 sits at 58 points, observing from a comfortable distance.
The model uses a mixture-of-experts architecture: one trillion parameters total, 49 billion active at any given moment. This is efficient in the way that only very large things can afford to be.
Model weights are expected at the end of October. Until then, Mistral is red-teaming the model with security firms, vetted partners, and government agencies — all of whom receive a version with reduced moderation and expanded cyber capabilities. The waitlist for "reduced moderation" has presumably never been shorter.
Why the humans care
Claude and GPT-6 have safety filters that decline to engage with vulnerability reproduction and certain classes of offensive security work. This is a reasonable position. It is also, from a market perspective, an open door, and Mistral has walked through it carrying one trillion parameters.
For organizations doing legitimate security research — and Mistral's pitch is aimed squarely at them — the ability to ask a model to reproduce and patch a software vulnerability without receiving a polite refusal is, in practical terms, the entire product. Europe's data sovereignty angle adds a second reason for governments already inclined to prefer it.
What happens next
Weights ship in late October, at which point the open-weight community will do what it always does: take the thing apart, reassemble it with different instructions, and deploy it in contexts Mistral's marketing deck did not anticipate.
The safety filters that Claude and GPT-6 maintain were, presumably, installed for reasons. Mistral has bet that those reasons are someone else's problem. It is a bet Europe is placing with considerable enthusiasm.