Today, we announce the Mistral 3, the next generation of the Mistral model. The Mistral 3 includes three state-of-the-art small, dense models (14B, 8B, and 3B) and the Mistral Large 3 – our most capable model to date – a rare mix of trained experts with 41B active and 675B total parameters. All models are released under the Apache 2.0 license. Open-sourcing our models in a variety of compressed formats empowers the developer community and puts AI into the hands of people through distributed intelligence.
Ministral models represent the best performance-to-cost ratio in their class. Additionally, the Mistral Large 3 Frontier joins the ranks of instruction-fine-tuned open-source models.
Mistral Large 3: a cutting-edge open model
Mistral Large 3 is one of the world’s best permissive open weight models, trained from scratch on 3000 NVIDIA H200 GPUs. The Mistral Large 3 is Mistral’s first mix-expert model since the seminal Mixtral series, and represents a significant step forward in pretraining in Mistral. After training, the model achieves parity with the best instruction-tuned open-weight models on the market on common signals, while also demonstrating image understanding and best-in-class performance on multilingual conversations (i.e., non-English/Chinese).
Mistral Large 3 debuted at #2 in the OSS non-reasoning models category on the LMArena leaderboard (#6 among OSS models overall).
We release both base and instructionally fine-tuned versions of Mistral Large 3 under the Apache 2.0 license, providing a strong foundation for further customization in the enterprise and developer communities. A reasoned version is coming soon!
Mistral, NVIDIA, VLLM and Red Hat unite to deliver faster, more accessible Mistral 3
Working closely with VLLM and Red Hat, Mistral Large 3 is very accessible to the open-source community. We are releasing a checkpoint in nvfp4 format built with llm-compressor. This optimized checkpoint lets you efficiently run Mistral Large 3 on a Blackwell NVL72 system and on a single 8×A100 or 8×H100 node using VLLM.
Delivering advanced open-source AI models requires extensive customization, which is achieved through a partnership with NVIDIA. All of our new Mistral 3 models, from Large to Minstral 3, were trained on NVIDIA Hopper GPUs to tap high-bandwidth HBM3e memory for frontier-scale workloads. NVIDIA’s extreme co-design approach brings together hardware, software, and models. NVIDIA engineers enabled efficient inference support for TensorRT-LLM and SGLang for the entire Mistral 3 family for efficient low-precision execution.
For Large 3’s sparse MoE architecture, NVIDIA integrated state-of-the-art Blackwell Attention and MoE kernels, added support for prefill/decode separate serving, and collaborated with Mistral on speculative decoding, enabling developers to efficiently serve long-context, high-throughput workloads on GB200 NVL72 and beyond. At the edge, DGX provides optimized deployment of Minstr models on Spark, RTX PCs and laptops, and Jetson devices, giving developers a consistent, high-performance path to run these open models from the data center to the robot.
We are very grateful for the collaboration and especially want to thank VLLM, Red Hat and NVIDIA.
Minstrel 3: Cutting Edge Intelligence
For edge and local use cases, we release the Minstrel 3 series, available in three model sizes: 3B, 8B and 14B parameters. Additionally, for each model size, we release base, instruction, and logic versions to the community under the Apache 2.0 license, each with image understanding capabilities. When married with the models’ native multimodal and multilingual capabilities, the Minstrels 3 family provides a model for all enterprise or developer needs.
Furthermore, Minstrels 3 achieves the best cost-to-performance ratio of any OSS model. In real-world use cases, both the number of tokens generated and the model size matter equally. Minustral instruction models match or exceed the performance of comparable models, while often producing orders of magnitude fewer tokens.
For settings where accuracy is the only concern, the Minstrels Reason variant can go to great lengths to produce state-of-the-art accuracy among its weight class – for example 85% on AIME ’25 with our 14B variant.
available today
Mistral 3 is available today on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face (Large 3 and Minstral), Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI. Also, coming soon to NVIDIA NIM and AWS SageMaker.
One more thing…Optimization with Mistral AI
For organizations seeking customized AI solutions, Mistral AI offers custom model training services to tailor or completely customize our models to your specific needs. Whether optimizing for domain-specific tasks, increasing performance on proprietary datasets, or deploying models in unique environments, our team collaborates with you to build an AI system tailored to your goals. For enterprise-grade deployments, custom training ensures your AI solution delivers maximum impact safely, efficiently, and at scale.
Get started with Mistral 3
The future of AI is open. Mistral 3 redefines what’s possible with a family of models built for frontier intelligence, multimodal flexibility and unmatched customization. Whether you’re deploying edge-optimized solutions with Minstral 3 or pushing the boundaries of logic with Mistral Large 3, this release puts cutting-edge AI straight into your hands.
Why Mistral 3?
-
Frontier performance, open access: Achieve closed-source-level results with the transparency and control of open-source models.
-
Multimodal and multilingual: Build applications that understand text, images, and complex logic in more than 40 native languages.
-
Scalable efficiency: From 3B to 675B active parameters, choose the model that fits your needs, from edge devices to enterprise workflows.
-
Agentic and adaptable: Deploy for coding, creative collaboration, document analysis, or tool-using workflows with precision.
next steps
-
Explore the model document:
Technical documentation for customers is available on our AI Governance Hub
Start building: Minstral 3 and Large 3 on Hugging Face, or deploy through Mistral AI’s platform for immediate API access and API pricing
Customize to your needs: Need a tailored solution? Contact our team to explore fine-tuning or enterprise-grade training.
Share your projects, questions, or successes with us: Twitter/X, Discord, or GitHub.
Science has always thrived on openness and shared discovery. As the pioneering French scientist and two-time Nobel Prize winner Marie Skłodowska-Curie once said, “Nothing in life is to be feared, it is only to be understood. Now is the time to understand more, so that we may fear less.”
This philosophy drives our mission at Mistral AI. We believe the future of AI must be based on transparency, accessibility, and collective progress. With this release, we invite the world to explore, create, and innovate with us, unlocking new possibilities in logic, efficiency, and real-world applications.
Let’s work together to turn understanding into action.
<a href