Mistral’s new 1T model aims to leapfrog closed and open rivals

Mistral is trying to plant a flag between two AI camps that increasingly look like they want different futures. On Tuesday, the French lab released Mistral Large 4, or ML4, a large multimodal model it believes can stand up to both American and Chinese systems. The company even gave it a nickname: Le Chonk. Very subtle, very on the nose.

For now, ML4 is not an open-weight release. Access runs through a public guardrail endpoint, and Mistral says the weights should arrive in three weeks, once safety testing is done. That delay is not just housekeeping. It sits right in the middle of the current anxiety around what open models can and cannot be used for, especially among enterprises and institutions that want auditability without inviting trouble.

Pierre Stock, Mistral’s VP Science, said the company will work with trusted partners and governments in the meantime so the eventual open-source weights can be used for defense, not malicious attacks. He also argued that open weights are easier to inspect, which is the kind of line that lands well with cautious buyers and regulators alike.

The engineering story is just as important. Mistral says ML4 was trained entirely on its own compute, using 4,000 Nvidia GPUs. Stock said that is two to three times fewer than the company’s Chinese competitors use, and significantly less than closed-source rivals. There are no benchmark results yet, so the bragging rights are still theoretical. But Mistral is betting that focused training will make ML4 strong in the places its customers care about most.

Those places include cybersecurity, finance, and chip design. That last one matters because chip design is central to two of Mistral’s biggest backers: ASML, which led its Series C, and Samsung, which led its Series D last month at a €21 billion valuation. Mistral has been careful not to look like an inference shop cosplaying as a frontier lab. ML4 is its argument that Europe can still be both useful and serious.