Mistral Large 3

Mistral Large 3 - Open-Weight Multimodal Mixture-of-Experts Model

Mistral Large 3 is Mistral AI's open-weight, general-purpose multimodal model built on a granular Mixture-of-Experts architecture. It pairs a 673B-parameter MoE language model with a 2.5B-parameter vision encoder, totaling 41B active and 675B total parameters, and is released under the Apache 2.0 license.

Architecture

• Granular MoE language model (673B total, 39B active) combined with a 2.5B-parameter vision encoder.
• Trained from scratch on 3,000 NVIDIA H200 GPUs.

Context & Deployment

• Supports a 256k-token context window.
• Available in FP8 (single 8xB200/H200 node) and NVFP4 (single 8xH100/A100 node) quantized formats, with BF16 for multi-node deployment.
• An EAGLE speculator checkpoint is provided for speculative decoding.

Multimodal & Tool Use

• Accepts image and text input; documented guidance recommends keeping images near a 1:1 aspect ratio.
• Offers native function calling and structured JSON outputs, intended for agentic use with a limited, well-defined toolset.