Mistral has released the weights of Mistral Large 4, the 1T-parameter multimodal flagship it trained from scratch on European infrastructure and previewed on October 6. After the preview, the model was fine-tuned with reinforcement learning for programming and cybersecurity tasks. The weights arrive as Mistral-Large-4-1T-A52B under a modified MIT license, and the model is reachable without local setup through the Mistral API, OpenRouter, Mistral’s Vibe Code environment and NVIDIA NIM containers.

On the Artificial Analysis Coding Agent Index the model scores 48% and spends 47% fewer tokens per task than GLM-5.3. On the Artificial Analysis Cyber benchmark it scores 51% at finding, reproducing and patching vulnerabilities, and Mistral says it audits code without security filters raising false positives. Three built-in reasoning modes, Low, Medium and High, trade cost against depth, with High as the default.

Official API pricing is $0.70 per million input tokens ($0.07 when cached) and $2.10 per million output tokens.

Related: Mistral Unveils Large 4, a 1T-Parameter Open-Weight Model Trained in Europe, Aleph Alpha Open-Sources Kolibri 1, a Bilingual German-English MoE Model

Mistral Large 4 open weights · Weights on Hugging Face