Mistral AI Releases Mistral 3 Family of Open-Weight Models

French AI company Mistral AI has launched the Mistral 3 family of open-weight models, announced in a press release. The release includes Mistral Large 3, a frontier-scale mixture-of-experts model, and nine smaller models under the Ministral 3 series.

Mistral Large 3 features 41 billion active parameters and 675 billion total parameters, with a 256,000-token context window. It supports multimodal and multilingual tasks and is released under the Apache 2.0 license. The model is trained on NVIDIA H200 GPUs and optimized for deployment across NVIDIA’s GB200 NVL72 systems and other platforms, offering high performance for enterprise workloads such as document analysis, coding, and content generation.

The Ministral 3 family comprises three model sizes—3B, 8B, and 14B parameters—each available in base, instruct, and reasoning variants. These models are designed for efficiency and local deployment, capable of running on a single GPU. They support vision inputs, long context windows, and multilingual processing, making them suitable for both enterprise and edge use cases.

All Mistral 3 models are available across multiple platforms, including Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, and IBM WatsonX, with support for NVIDIA NIM and AWS SageMaker coming soon. The models can be fine-tuned or customized for domain-specific tasks through Mistral AI’s enterprise services.