Mistral AI Launches Mistral Large 4 ('Le Chonk') Flagship Model
Mistral AI has launched the public preview of Mistral Large 4 (codenamed 'Le Chonk'), a 1-trillion-parameter natively multimodal model accessible via API. Weights for the open-source model are scheduled for public release by the end of the month following red-teaming safety evaluations. As a massive open-weight frontier model matching or outperforming top open models in enterprise workloads like cybersecurity, law, and finance, Mistral Large 4 offers organizations high-performance AI without relying on closed systems. It reinforces model diversity and cost competition against proprietary frontier AI offerings. Mistral Large 4 utilizes a Mixture-of-Experts (MoE) architecture with 1 trillion total parameters and 49 billion active parameters per token. It was trained from scratch in Europe using 3,800 NVIDIA Grace Blackwell GPUs in strict compliance with EU regulatory standards.
## BACKGROUND
Frontier models refer to the most capable general-purpose AI systems setting the state of the art in machine learning performance. Mistral AI is a French laboratory known for pioneering high-performing open-weight models using Mixture-of-Experts (MoE) architectures, which route inputs to specialized sub-networks to achieve high capabilities with greater inference efficiency.