Mistral AI Introduces Mistral 3 Model Family, Including Mistral Large 3

Mistral AI has announced the release of Mistral 3, a new generation of artificial intelligence models. The family includes three dense Ministral 3 models with parameter sizes of 3B, 8B, and 14B, as well as the flagship sparse mixture-of-experts model Mistral Large 3, trained on 3,000 NVIDIA H200 GPUs, with 41 billion active parameters out of 675 billion total. All models are distributed under the Apache 2.0 license. Mistral Large 3 is the first model in the family since the Mixtral series to use a mixture-of-experts architecture. After post-training, the model has achieved parity with the best open models that have undergone instruction fine-tuning, and demonstrates image understanding as well as high performance in multilingual conversations. The model ranks second in the non-reasoning OSS model category on LMArena. Mistral Large 3 is released in both base and instruct versions, with a reasoning version to follow later. In collaboration with NVIDIA, Red Hat, and vLLM, inference has been optimized on Blackwell NVL72 hardware and on a single node with 8×A100 or 8×H100. Ministral 3 models are available in base, instruct, and reasoning variants, all with support for images and multilingual capabilities. The models support efficient operation on edge devices, including DGX Spark, RTX PC, and Jetson. Mistral 3 is already available on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI. Customization services for organizational needs have also been announced.