The recently released weights for Mistral’s flagship 1.05 trillion-parameter AI model, codenamed Le Chonk, place it ahead of all other non-Chinese open-source models on the Artificial Analysis Intelligence Index. The model scored 26 on the index, one point above the U.S.-developed Muse Spark Open, while it ranked fifth among all open-weight models on the Vals Index. Mistral used 3,800 Nvidia Grace Blackwell GPUs to train the natively multimodal Mixture-of-Experts model, which features 52B active parameters and a 1M token context window.
Mistral cut the cost of using the model by about 50%, pricing it at $0.70 per 1M input tokens and $2.10 per 1M output tokens compared to the previous preview pricing of $1.36 and $4.18, respectively. The new version reduces hallucinations to 28% from 41% in the preview and uses 39% fewer output tokens per Intelligence Index task. The weights are available under a MEOW license, which requires a separate commercial agreement for businesses operating the model as a service.
