Mistral AI announced the release of Mistral 3 on December 2, 2025, introducing a new series of open-source models. The release includes three small, dense models (14B, 8B, and 3B parameters) and Mistral Large 3, a sparse mixture-of-experts model trained with 41B active and 675B total parameters. All models are released under the Apache 2.0 license, providing developers with access to advanced AI capabilities. Mistral Large 3 is one of the best permissive open weight models, trained from scratch on 3000 NVIDIA H200 GPUs. It represents Mistral’s first mixture-of-experts model since the Mixtral series and achieves parity with the best instruction-tuned open-weight models on general prompts. The model also demonstrates image understanding and best-in-class performance on multilingual conversations. Mistral Large 3 debuted at #2 in the OSS non-reasoning models category on the LMArena leaderboard. The base and instruction fine-tuned versions of Mistral Large 3 are available under the Apache 2.0 license, offering a foundation for further customization. A reasoning version is expected soon. Mistral collaborated with NVIDIA, vLLM, and Red Hat to optimize Mistral 3 for accessibility. The models were trained on NVIDIA Hopper GPUs using HBM3e memory for high-bandwidth workloads. NVIDIA’s TensorRT-LLM and SGLang support enable efficient low-precision execution. For Large 3’s sparse MoE architecture, NVIDIA integrated Blackwell attention and MoE kernels, supporting prefill/decode disaggregated serving and speculative decoding. Ministral 3 models are optimized for edge deployment on DGX Spark, RTX PCs, laptops, and Jetson devices. The series includes 3B, 8B, and 14B parameter models with base, instruct, and reasoning variants, all under Apache 2.0 license. Ministral 3 offers the best cost-to-performance ratio among OSS models. In real-world use cases, Ministral instruct models match or exceed comparable models while generating fewer tokens. The reasoning variants can achieve state-of-the-art accuracy, such as 85% on AIME ‘25 with the 14B variant. Mistral 3 is available on multiple platforms including Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, and Together AI. Custom model training services are available for organizations seeking tailored AI solutions. Mistral AI provides enterprise-grade deployments for secure, efficient, and scalable AI systems. The release emphasizes open access, multimodal flexibility, and customization for various applications. *Source: [mistral](https://mistral.ai/news/mistral-3/)*