According to CoinMeta, Mistral has released a new flagship model, Mistral Large 4, which adopts a MOE architecture with a total of 1.05 trillion parameters, but only 49 billion parameters are activated during each inference process. The model natively supports image and text input, with a context length of up to 1 million token. A preview of API is already available, and the model weights are scheduled to be made public on October 27th. Mistral claims that Large 4 is currently the strongest open-weight model developed in the United States and Europe in terms of comprehensive performance. The official Deepswe software engineering evaluation scores v1.1 at 62%, the financial evaluation Finch at 67%, and the satellite image target positioning Dior-RSVG at 73%. However, these results mainly come from Mistral self-tests and are not based on a unified standard. Large 4 was trained from scratch and used approximately 4000 NVIDIA Grace Blackwell GPU to train for about two months in Europe's own data center. Mistral also highlights "European autonomy" as a selling point; from training to API services and future autonomous deployment, the model can remain on European infrastructure.