Mistral releases the Mistral 3 open-weight model family

Mistral AI/Mistral models

Mistral released the Mistral 3 family under Apache 2.0: Mistral Large 3 with 675B total and 41B active parameters trained on 3,000 NVIDIA H200 GPUs, and three smaller Ministral models (3B, 8B, 14B). All handle images and text across 40+ languages.

• Mistral Large 3 ranks second on LMArena for open-source non-reasoning models and excels at multilingual tasks.
• The 14B reasoning variant scores 85% on AIME 2025, competitive with frontier models.
• Available through Mistral AI Studio, Amazon Bedrock, Azure, Hugging Face, and other platforms.