One trillion parameters, trained on 4,000 Nvidia GPUs. Those are the two numbers Mistral AI attached to Mistral Large 4, the model it released on Tuesday and nicknamed Le Chonk.
The Paris lab says that chip count is two to three times lower than its Chinese competitors use and far below the closed American labs, TechCrunch reported, citing Pierre Stock, its vice-president of science.
The model is tuned for cybersecurity, finance and chip design. Mistral’s backers have an interest in at least one of those. ASML led its Series C, and Samsung led a Series D last month that valued the company at €21 billion.
For now the model is reachable only through a hosted endpoint with guardrails. Mistral says the weights will follow in about three weeks, once safety testing is finished.
Stock said the company would work with trusted partners and governments so that the open weights are used for defence and not for malicious attacks. That is an awkward promise. Once weights are published, the publisher has little say over who runs them or what the guardrails are replaced with.
Guillaume Lample, the chief scientist, told Decrypt it is by far the strongest open-weight model from the US or Europe. The wording leaves out China’s open models and every closed one.
Neither claim can be checked yet. TechCrunch notes that benchmark results are still pending, and with access limited to Mistral’s own endpoint nobody outside can test the efficiency figure either.
The date to watch is the end of October. Either the weights arrive on schedule, or safety testing finds something and Europe’s best-funded open-model lab has to explain a delay. Mistral has not said what result would hold the release back.
