The easy reading of today’s Mistral Large 4 launch is that Europe is back at the frontier, and I would like that to be true. It is a 1 trillion parameter model with 49 billion active, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters, and for now it is an API preview with open weights promised by the end of the month.

Mistral presents it to Wired as the best open-weight model outside China and “very, very close” to some proprietary models, which by its own account leaves some closed models still ahead.

Whether Europe is really back will show at the end of October, in how many teams run those weights on their own infrastructure.