Mistral says ML4 was trained using 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe and much of its training data was multilingual (Mistral Blog)

Mistral has unveiled the public preview of its latest model, Mistral Large 4, or ML4, which demonstrates exceptional performance across various domains. This 1 trillion-parameter model is Mistral's most capable to date and will be released by the end of the month. The model was trained using 3,800 NVIDIA Grace Blackwell GPUs housed in Mistral's European datacenters.

ML4 is a natively multimodal model with 49 billion active parameters and exhibits exceptional performance in coding, agentic workflows, and multimodal understanding, even surpassing frontier closed models in some areas such as visual grounding. The model's multilingual training data, covering over 160 languages, including all EU official languages, is a notable feature.

ML4 has been trained with the same methodology used for customers via Mistral Forge, and it will be available in multiple regions worldwide. The model's capabilities extend beyond cybersecurity, as it has shown proficiency in software engineering, repository understanding, and complex terminal workflows. ML4 scored second in the AutomationBench, which evaluates complex workflows across various applications, and it ranks among the top five models globally in the Artificial Analysis Cyber Index, leading open-weight models outside China.