Mistral has released a preview of Mistral Large 4, a one trillion parameter model running on 3,800 NVIDIA Grace Blackwell GPUs. The system is currently accessible via their API with open weights scheduled for release later this month. Testing the model on image generation tasks reveals two distinct reasoning modes: a standard setting and a high reasoning setting. A visual comparison of pelican images shows the high reasoning mode produces a cleaner result with correct anatomical details, yet it generates fewer output tokens than the standard mode.
Artificial Analysis scores the model at 38, placing it just behind DeepSeek 4.1 Flash while representing a significant jump from the previous version. The earlier Mistral Large 3 scored only 9 on the same benchmark and struggled to render coherent imagery. This new release brings Mistral back to a position roughly six months behind the current frontier.
* Model size: 1 trillion total parameters with 49 billion active
* Hardware: Trained on a cluster of 3,800 NVIDIA Grace Blackwell GPUs
* Benchmark: Artificial Analysis score of 38
