Mistral Large 3 is a general-purpose, multimodal model designed to serve as a highly capable foundation for diverse applications. Built as a sparse mixture-of-experts, the architecture utilizes 41 billion active parameters out of a 675 billion total parameter count. This design allows the model to balance depth and efficiency, positioning it as a versatile tool for users who require a robust, state-of-the-art system capable of handling complex, multi-faceted inputs and generating high-quality text outputs.

The model was trained from scratch using a massive compute infrastructure consisting of 3,000 NVIDIA H200 GPUs, reflecting a significant investment in its foundational development. As a frontier instruction-fine-tuned model, it is engineered to excel in demanding production environments. By combining this intensive training lineage with a flexible, multimodal design, the model provides a strong, forward-looking solution for developers and enterprises seeking to integrate advanced, high-performance intelligence into their workflows.