Mistral Large 4 Preview

Mistral Large 4 Preview (also known by its internal moniker "le Chonk") is a flagship natively multimodal language model developed by Mistral AI. Designed to push the frontier of open-weight performance, it features an advanced granular Mixture-of-Experts (MoE) architecture totaling approximately 1.05 trillion parameters with 52 billion active parameters per token, alongside a separate vision encoder. The model was trained from scratch on thousands of NVIDIA Grace Blackwell GPUs within Europe and supports rich multilingual and multi-step agentic workflows.

Capabilities and Performance

Mistral Large 4 demonstrates state-of-the-art performance among open models across critical enterprise verticals such as coding, legal reasoning, finance, and cybersecurity. Its native multimodality provides enhanced visual grounding and document reasoning capabilities. The public preview allows developers to evaluate its advanced tool use and complex logical inference via API before the full open-weight model distribution.