mistral-large-2512

Mistral Large 3 is a general-purpose multimodal model built on a granular mixture-of-experts architecture. It utilizes 675B total parameters, with 41B active parameters engaged during inference to balance performance and efficiency. Designed to handle diverse and demanding workloads, the model supports a 256k token context window and native vision capabilities, making it well-suited for complex reasoning, document analysis, and large-scale data processing.

The model was trained from the ground up using a cluster of 3000 H200 GPUs, reflecting a significant investment in computational scale. Following its initial training, the model underwent instruction-based post-training to refine its performance for chat, agentic workflows, and specialized instruction-following tasks. Available in various precision formats, including FP8 and BF16, it is engineered for production-grade reliability, allowing for deployment in both cloud environments and on-premises infrastructure for enterprise-level applications.