Mistral Large 4 released: 1.05 trillion parameters, claims to have made the strongest open-source model in Europe and America.

动察 Beating AI News Flash: French AI company Mistral has released its new flagship model Mistral Large 4, internally nicknamed "Le Chonk." It uses a MoE architecture with a total of 1.05 trillion parameters, but only activates 49 billion parameters per inference. The model natively supports image and text input, with a maximum context of 1 million tokens. The API is already open for preview, and the model weights are planned to be made public on October 27.

Mistral claims that Large 4 is currently the strongest open-weight model developed in the United States and Europe in terms of overall performance. The official DeepSWE v1.1 software engineering benchmark score is 62%. In its comparison table, GLM-5.3 is 61%, DeepSeek V4 Pro is 57%, and Qwen 3.8 Max is 51%. On the Finch financial benchmark, it scored 67%, tied with DeepSeek V4 Pro; on DIOR-RSVG satellite image target localization, it scored 73%, higher than GPT-6 Astra's 68%.

However, these results currently mainly come from Mistral's own testing and are not based on a unified standard. When Zhipu released GLM-5.3, the score it announced on the same DeepSWE v1.1 was 66.9%, while Kimi K3 reached 67.5%, both higher than Large 4's 62%. Different teams may use different test configurations and run methods.

Large 4 was trained from scratch. Mistral used about 4,000 Nvidia Grace Blackwell GPUs and trained for about two months in its own data center in Europe. Mistral also uses "European autonomy" as a selling point: from training and API services to future autonomous deployment, everything can remain on European infrastructure.