Mistral Large 3

Mistral Large 3 is a multimodal language model announced by Mistral AI on December 2, 2025, as part of the Mistral 3 family. The model uses a sparse Mixture-of-Experts (MoE) architecture with 41B active and 675B total parameters, and a 256k context window.

It was trained from scratch on 3,000 NVIDIA H200 GPUs and is Mistral AI’s first MoE model since the Mixtral series. Both base and instruction-fine-tuned versions are released under the Apache 2.0 license. The model supports image understanding (multimodal input) and multilingual conversations.

Mistral Large 3 is available through Mistral AI Studio, Amazon Bedrock, Azure AI Foundry, Hugging Face, IBM WatsonX, OpenRouter, Fireworks, and Together AI. The model weights can run on a single 8×A100 or 8×H100 node with vLLM, and in NVFP4 format on Blackwell NVL72 systems.