Introducing Mistral Large 4

Mistral AI today announced a public preview of Mistral Large 4, codenamed ML4, positioning it as the company’s largest and most capable open-weight model to date. The model features 1 trillion total parameters with 49 billion active parameters and is natively multimodal, supporting text, vision, and agentic workflows. A preview API is available immediately on Mistral Studio, with full model weights scheduled for release by the end of the month. The model demonstrates strong performance across coding, cybersecurity, finance, and law, often surpassing other open-weight models developed in the US or Europe.

ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters, leveraging infrastructure designed for high-performance AI workloads. The model’s training data spans over 160 languages, including every official EU language, reflecting its multilingual capabilities. Mistral has collaborated with leading enterprises in finance, manufacturing, and public sector roles to refine ML4, using the same training and customization pipeline offered through Mistral Forge to enterprise customers.

The model excels in cybersecurity applications, ranking among the top five globally on the Artificial Analysis Cyber Index and leading open-weight models outside China. It scores 82% on a test reproducing and patching real software vulnerabilities, outperforming several closed models that refuse such tasks due to safety filters. ML4 also achieves 93% on Cybench’s 40 security challenges, demonstrating practical utility in vulnerability analysis, malware assessment, and detection rule generation. Its cyber capabilities are paired with open weights and self-deployment options, enabling organizations to run advanced security work under their own policies.

ML4 introduces significant advances in multimodal and scientific reasoning, combining vision with agentic workflows for tasks like inspecting satellite imagery or verifying engineering drawings. In scientific domains, it leads open-weight models on SciCode-Verified benchmarks and can execute complex simulations such as Hartree–Fock calculations in a single step. The model also performs strongly in coding and knowledge work, scoring 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase, while ranking second in a blind human evaluation of coding quality among five models. A European deployment ensures compliance with local regulations and independent operation.