Mistral Large 4 enters public preview: 1T-parameter multimodal MoE, weights promised by end of October
In one sentence On October 6 Mistral opened the public API preview of Mistral Large 4, a natively multimodal MoE with 1 trillion total and 52 billion active parameters, trained from scratch on 3,800 Grace Blackwell GPUs in Europe. Weights are announced for the end of the month.
Mistral, Europe's leading AI lab, showed its new flagship: Mistral Large 4. For now it is available as an API preview on the Mistral Studio platform; the weights, meaning the downloadable model that can be installed on your own servers, are promised for the end of October.
It is a very large model, one trillion parameters, but it activates about 52 billion per response thanks to its mixture-of-experts architecture. It understands text and images and combines fast-answer and reasoning modes in a single model. It was trained entirely in European data centers and covers more than 160 languages, including all official EU languages.
Mistral highlights cybersecurity, finance and law as its strongest areas, and on-premise installation for those who cannot send data outside. Benchmark results are company-reported and reinforcement learning is still running, so the numbers may change before the weights ship.
Companies
Mistral AI
Tools
Mistral Large 4, Mistral Studio
Tags
Sources