Mistral Large 4

Today, Mistral announced the public preview launch of their new AI model, Mistral Large 4, also unofficially known as ML4. ML4 sets a new standard for open-weight performance and pushes the boundaries of what's possible with open models. The model is a 1 trillion-parameter, natively multimodal system with 49 billion active parameters, making it their largest and most capable model to date.

ML4 demonstrates exceptional performance across various domains, including coding, agentic workflows, and multimodal understanding. It already matches the performance of the strongest open-source models globally while outperforming any open-weight model developed in the US or Europe. In critical enterprise workloads, such as cybersecurity, finance, and law, ML4 stands out as state-of-the-art among open models. In some areas like visual grounding, it even surpasses frontier closed models.

The model was trained using 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters and will be available across multiple regions worldwide, including a European deployment operated end-to-end, independent of other digital service providers and under European law. ML4 was trained on multilingual data, covering over 160 languages, including all official EU languages.

Mistral has been collaborating with leading enterprises in various industries, including finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping, public sector, and more, to train ML4.

The model is still in the early stages of development, and more details on its architecture, additional benchmarks, and post-training methodology will be shared as the release progresses. ML4 will serve as the foundation for a new generation of specialized and optimized Mistral models. The model's strong performance in cybersecurity is particularly noteworthy, as it ranks among the top five models globally on the Artificial Analysis Cyber Index and leads open-weight models developed outside China.

In one test, ML4 scored 82%, the highest among all models. Its capabilities extend beyond what it was explicitly trained for, as evidenced by internal testing where ML4 proved useful for analyzing malware, prioritizing vulnerabilities, and writing detection rules.