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 Large 4's open weights mean you can deploy a cutting-edge multimodal AI model independently, without vendor lock-in.

Mistral AI has launched a public preview of its new AI model, Mistral Large 4, also referred to as ML4 or "le Chonk." This 1 trillion-parameter, natively multimodal model boasts 49 billion active parameters and is positioned as Mistral's most capable model to date. ML4 demonstrates strong performance in coding, agentic workflows, and multimodal understanding, claiming to be competitive with leading open-source models and outperforming many US and European developed open-weight models, particularly in enterprise workloads like cybersecurity, finance, and law. It also shows promise in areas like visual grounding, surpassing some closed frontier models. Mistral trained ML4 from scratch using 3,800 NVIDIA Grace Blackwell GPUs housed in its own European data centers, underscoring its commitment to AI sovereignty. The model's training data included a significant portion in over 160 languages, including all official EU languages. The weights for ML4 are slated for release by the end of the month, following a period of real-world red-teaming with security experts and authorities. Mistral plans to release further details on the model's architecture and benchmarks.

Mistral's announcement of ML4, an open-weight model trained on European infrastructure, signals a significant push for AI sovereignty, aiming to provide businesses with greater control over their AI deployments. The model's multilingual training data suggests a focus on global applicability and a potential advantage in localized AI solutions. By releasing open weights, Mistral directly challenges the dominance of proprietary models and offers a compelling alternative for organizations prioritizing autonomy and data governance, especially in sensitive sectors.

The performance claims of ML4, particularly in cybersecurity and multimodal tasks, position it as a direct competitor to leading closed-source models, potentially lowering the barrier to entry for advanced AI capabilities. This development could accelerate innovation by making state-of-the-art AI more accessible and customizable for a wider range of businesses and researchers. The emphasis on self-deployment and European data centers also highlights a growing trend towards regionalized AI ecosystems, driven by both technological advancements and regulatory considerations.