Mistral AI announced Mistral Large 4 on October 6, 2026, a natively multimodal mixture-of-experts model with 1 trillion total parameters and 52 billion active per token. A public preview API went live on Mistral Studio at $1.36 per million input tokens and $4.18 per million output tokens. Mistral said the weights would be released “by the end of the month,” with more detail on architecture and post-training to follow.
The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centers, and Mistral says training is still ongoing with “no signs of saturation.” Its reinforcement-learning setup runs on about 3,000 GPUs and produces roughly 33 billion tokens a day.
Mistral’s reported scores include 61.7 percent on DeepSWE v1.1, 28.3 percent on Terminal-Bench 4, 82 percent on a reproduce-and-patch vulnerability test and 93 percent on Cybench. It claims a top-five place on the Artificial Analysis Cyber Index and a wide lead over other non-Chinese open-weight models, and says the model is more willing than closed models to perform security tasks those models refuse.
Why it matters: it is the largest model Mistral has announced and the clearest European attempt at a frontier-scale open-weight model, trained on European-owned compute, at a time when the leading open-weight releases have mostly come from China.
What it does not show: at announcement the weights were not yet out and no license was stated, the benchmark figures are Mistral’s or its partners’, and a human coding evaluation in the same post placed it second of five, behind Claude Opus 5.