Mistral Drops a 1-Trillion-Parameter Fat Cat: Europe Catches Up to Chinese Open Models, Charges Three Weeks of API Fees Before Releasing the Weights
On October 6 France's Mistral launched a public preview of Mistral Large 4, code-named Le Chonk. At 1 trillion total parameters it is the largest model the company has built, and the weights are scheduled for release at the end of October.
Three key facts
The architecture is sparse mixture-of-experts: 1 trillion parameters in total, 49 billion active per inference. It takes multimodal input and produces text, blends instruct and reasoning modes, and covers more than 160 languages including every official EU language. Mistral says training ran on 3,800 Nvidia Grace Blackwell GPUs entirely in its own European data centers; VentureBeat reports the run took about two months. API pricing is $1.36 per million input tokens and $4.18 per million output tokens.
Nearly every rival in the official comparison is a Chinese open-weight model. On the DeepSWE v1.1 coding benchmark it scores 61.7%, against 61% for GLM-5.3 and 57% for DeepSeek V4 Pro in VentureBeat's tally. On Dense200 visual grounding it posts 42%, edging GPT-6 Astra's 41%. On the Cybench security test it hits 93%, which Mistral calls the highest open-weight score. In a blind human evaluation run by Surge AI, though, it lands second at 3.74 out of 5, behind Claude Opus 5 at 4.22. Co-founder and chief scientist Guillaume Lample's line: ML4 is at the frontier of open weight models.
The rollout is API first, weights later. The preview is available now as a paid service through Mistral Studio. Weights follow after a roughly three-week test window under a custom Mistral license. Cybersecurity leaders, vetted partners and state authorities get a red-team build with reduced moderation and expanded cyber capabilities. The backdrop is September's 3 billion euro Series D at a valuation above 21 billion euros.
WangDou's Take
One trillion parameters sounds like an aircraft carrier launch. Then you read the fine print: only 49 billion show up for work on any given call. That is a 4.9% usable-floor-space ratio, and the other 95% is hallway. How did the sign that says frontier of open models get hung? Coding at 61.7% versus GLM-5.3 at 61% is a 0.7-point lead. That is not overtaking, that is standing side by side for a group photo. Put real humans in a blind test and it is 3.74 against Claude's 4.22, so the press release only lines it up against Chinese open models, and the one closed-model matchup it picks is a vision task it wins by a single point. Then the schedule: the word open is in the headline, the weights arrive at month's end, and the first three weeks are billed at $1.36 a million tokens. Is that open source, or a pre-order? Europe's 3 billion euros of sovereign AI bought a fat cat that caught up to the Chinese team. Caught up. Did not pass.
Source: Mistral AI · Introducing Mistral Large 4 · VentureBeat