Mistral 3 family of models released

The new generation of models includes three small, dense models (3B, 8B, and 14B parameters) and a sparse mixture-of-experts model, Mistral Large 3 (41B active parameters). All models are released under the Apache 2.0 license, promoting open access.

Mistral Large 3 is presented as a state-of-the-art open model, achieving parity with leading instruction-tuned open-weight models and demonstrating strong multilingual conversation capabilities. The smaller Ministral models offer the best performance-to-cost ratio in their class, suitable for edge and local use cases, and include base, instruct, and reasoning variants with image understanding.

Accessibility and efficient deployment are key features, supported by collaborations with NVIDIA, vLLM, and Red Hat, optimizing the models for various NVIDIA hardware configurations. The models are available across numerous platforms, and custom training services are offered for tailored enterprise solutions. The release emphasizes frontier performance, multimodal flexibility, and scalability from edge devices to data centers.