Mistral Large 4 preview runs on 3,800 GPUs

Mistral says Large 4, its new 1-trillion-parameter model with 49 billion active parameters, was trained on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers, and the company is hosting the current preview on its API. The API is available now, while Mistral plans to publish the model weights on October 27, so open-weight teams cannot yet self-host it. Builders comparing the launch with the doubled open-weight coalition should record Mistral’s scale and test cost and performance on their own workloads before relying on company benchmark claims.