Mistral AI Unveils Powerful Mistral 3 Model Family
- •Mistral AI has launched the Mistral 3 family featuring the high-capacity Mistral Large 3 and edge-optimized Ministral series.
- •Mistral Large 3 utilizes a sparse Mixture-of-Experts architecture with 675 billion total parameters to achieve elite reasoning performance.
- •All models are released under the Apache 2.0 license and support native multimodal capabilities across over 40 languages.
Mistral AI has introduced its next generation of open-source models, the Mistral 3 family, designed to deliver high performance through diverse model scales. The flagship, Mistral Large 3, employs a sparse Mixture-of-Experts architecture that activates 41 billion of its 675 billion total parameters for any given task. This design enhances efficiency while maintaining top-tier performance on global benchmarks. The model competes directly with leading proprietary systems, offering a robust alternative for enterprise-level deployment.
For localized use cases, the company released the Ministral series with 3B, 8B, and 14B parameters optimized for edge computing. These models process data directly on devices like laptops, ensuring privacy and reducing latency. Notably, the 14B reasoning variant demonstrated exceptional logical capabilities by scoring 85% on the AIME '25 mathematics benchmark. These compact versions are designed to bridge the gap between low-resource environments and complex reasoning tasks.
To optimize hardware compatibility, Mistral AI partnered with NVIDIA to implement the NVFP4 numerical format, allowing Large 3 to run effectively on standard enterprise infrastructure. All models in this release use the Apache 2.0 license, promoting transparency and open modification. Beyond text, the models are natively multimodal and support over 40 languages to serve a global developer base. Mistral also provides fine-tuning services to align these AI tools with specific business objectives and proprietary data.