Gangsta AI
A French Startup Just Dropped a Trillion-Parameter Open Model and Named It “Le Chonk”
Open Models
By Keanu Reeves · 2026-10-06 · 4 min read
They named it le Chonk.
That is not a typo, and it is not a leak. It is the official-unofficial nickname French startup Mistral gave the model it unveiled today: Mistral Large 4, a 1-trillion-parameter, natively multimodal system that the company calls its largest and most capable to date. A public preview is live now; the actual weights, Mistral says, drop at the end of this month.
The number everyone will repeat is the trillion. The number that matters more is the other one: 49 billion active parameters. ML4 is a mixture-of-experts model, so it carries a trillion parameters of knowledge but only fires a slice of them per token — big brain, light footprint. Mistral says it is competitive with the strongest open-source models anywhere on Earth, and that it “significantly outperforms any open-weight model developed in the US or Europe.” On enterprise-flavored work — coding, cybersecurity, finance, law — the company claims state-of-the-art among open models, and on some vision tasks says it edges past even frontier *closed* models.
“A trillion-parameter model you can eventually download, run, and inspect yourself — from a company that is not in California or Beijing.”
Forged in Europe, by design
Mistral is leaning hard into the sovereignty story. ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview is served off that same iron. The pitch is not subtle: open weights, European infrastructure, and — their words — “control over your AI.” For a continent that spent the last two years watching American and Chinese labs trade the crown, an open model that claims the best-outside-China title is a genuine flag in the ground.
The underdog framing writes itself. A startup, not a trillion-dollar incumbent, shipping something this heavy and then *giving the weights away*. If the benchmarks hold once the weights are public and independent testers get their hands on le Chonk, this is the kind of release that nudges the whole open ecosystem forward.
But “best” is a moving target
Here is the honest part. “Beats every US and European open model” is Mistral's own claim, measured by Mistral, on a preview that outsiders cannot yet fully probe. Last week Google called Gemini 4 the most powerful model on the planet. The week before that it was someone else. Every lab's launch post declares itself the new leader, and every one of them is telling a version of the truth that happens to flatter them.
So when a headline hands you a single model and a single superlative, the move is not to believe it or dismiss it — it is to check it against the others. Ask the same hard question of Mistral, ChatGPT, Claude, Gemini, and Grok side by side, and the disagreements are where the truth leaks out. One model's confident answer is a guess; a dozen frontier models cross-examined is something closer to a verdict. That is the entire premise of asking every top model at once and fusing the replies into one cited answer instead of betting everything on whoever shipped most recently.
le Chonk might be exactly as good as Mistral says. The only way any of us will actually know is by comparing — which, conveniently, is the one habit worth keeping no matter who's winning this week.
Curious which model actually wins your question? See how the top models stack up →
Sources / Receipts
More: Best AI models · Compare all AI · Frontier Models · All articles