France's Mistral previews a giant open AI model built in Europe
Paris-based Mistral on Tuesday opened a preview of Mistral Large 4, a 1-trillion-parameter AI model, through its API. The company says it is the most capable open-weight model built outside China. Anyone will be able to download and run it on their own servers once Mistral releases the files, which it says will happen by month's end.
Europe has a new heavyweight AI model. Mistral, a French startup, on Tuesday opened a preview of Mistral Large 4, a model with about 1 trillion parameters, the settings a model learns in training. Developers can try it now through Mistral's API.
The bigger moment comes later. Mistral says it will release the model's weights, the files that let anyone download and run it on their own machines, by the end of October. The Next Web put the date at Oct. 27. Until then, Mistral is testing it with cybersecurity firms, vetted partners and government agencies, who get a version with fewer limits.
Mistral calls Large 4 the strongest open-weight model built outside China. It uses only 49 billion of its parameters for any one task. It handles text and images and was trained on more than 160 languages, including every official EU language.
Security is the pitch. Mistral says leading closed models, including Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra, score near zero on one test because they refuse to reproduce a real software flaw. Defenders often need to prove a flaw is real before fixing it. "The cyber defense capabilities will enable enterprises and governments to defend themselves," said co-founder Guillaume Lample, as quoted by The Next Web.
Early scores look strong but are not final. On one coding test, Large 4 scored about 62%, edging China's GLM-5.3 at 61%. Mistral says training is still under way and results may change.
The model was trained from scratch on Nvidia chips in Mistral's own European data centers. Mistral says it used 3,800 of them; The Next Web said about 4,000. It is the first model funded by Mistral's 3 billion euro round in September. "The narrative that Europe cannot compete is something that is not true," CEO Arthur Mensch said.
What we know: Large 4 is in preview now, with downloadable files promised this month.
What we don't: How it scores once training finishes and outsiders can test it.
Where outlets agree Mistral and The Next Web both say Large 4 has about 1 trillion parameters, uses 49 billion per task, and will have downloadable weights later this month.
Where they differ Mistral says it trained the model on 3,800 Nvidia GPUs, while The Next Web reported about 4,000, and all benchmark scores are still preliminary.