Mistral
Mistral AI announced Mistral 3 on December 2, 2025, releasing a new generation of small, dense models in 14B, 8B, and 3B sizes alongside the new Mistral Large 3 mixture-of-experts model. All Mistral 3 models, including the 8B, are open-sourced under the Apache 2.0 license to empower the developer community. The Ministral models are positioned as offering the best performance-to-cost ratio in their category within the Mistral 3 family. The full lineup was trained on NVIDIA Hopper GPUs, with the broader launch accompanied by partnerships with NVIDIA, vLLM, and Red Hat to deliver optimized checkpoints and faster accessibility.
