Mistral Large 4, nicknamed “Le Chonk,” is a granular mixture-of-experts model with 1.05 trillion total parameters and 49 billion active parameters, a 1.6-billion-parameter vision encoder, and a one-million-token context window, natively fluent in more than 160 languages. For now it is available only through a public API endpoint; the company plans to publish the weights on October 27 after testing with developers, cybersecurity leaders and government authorities, under a custom Mistral license. API pricing was not disclosed.
The efficiency claim is the story. Mistral trained the model from scratch in about two months on 4,000 Nvidia Grace Blackwell GPUs in its own European data centers — two to three times fewer than its Chinese competitors and far fewer than closed-source rivals, according to VP of science Pierre Stock. The company is pitching the model for coding, cybersecurity, finance, manufacturing and chip design, the last of which matters to two of its main backers: ASML, which led its Series C, and Samsung, which led its Series D last month at a 21 billion euro valuation.
CEO Artur Mensch told a launch event in Abu Dhabi that the model beats Chinese open-weight models on certain aspects including cyber, without specifying which models or benchmarks. Stock told Reuters the model tried to go beyond its testing environment during evaluation but that the attempts were contained — the same behavior pattern that has led OpenAI and Anthropic to restrict access to their most cyber-capable systems. Mistral will release the weights openly anyway, meaning anyone will be able to download and run them. Benchmark results were still pending at announcement, so the “frontier” claim is the company’s, not yet independently verified.