Mistral Large 4 tops non-US, non-China models but lags Chinese open rivals
Mistral introduced its trillion-parameter Large 4 model, which Artificial Analysis ranked as the strongest AI system from outside the United States and China. It scored 38 on the Intelligence Index and 50 on the Cyber Index, matching GLM-5.3-Flash but trailing MiMo-V2.6-Pro on cyber capability. Several Chinese open-weight models still outperformed it overall, and Mistral plans to release the weights by the end of October.
Mistral's Large 4 is a mixture-of-experts system with roughly one trillion parameters, 49 billion active, trained on Nvidia Grace Blackwell hardware. Artificial Analysis placed it at 38 on its intelligence ranking and 50 on its cyber ranking, while its CyberGym-E2E-AA result reached 82%.
The model supports 512,000-token context and accepts text plus images, with text output. API pricing is $1.36/$4.18 per million input/output tokens, or half that for two weeks; cached input costs $0.14 per million. Mistral intends to publish weights by late October.
Mistral's showing may affect European developers and enterprises seeking alternatives to US and Chinese systems. Its open-weight release could let smaller teams inspect, adapt, and deploy the model locally, though higher task costs than some rivals may constrain widespread use. Strong cyber benchmark results could draw interest from security researchers, while benchmark limitations mean real-world impact remains uncertain.