Mistral AI has announced the launch of Mistral 3, a next-generation model family, including Mistral Large 3 for a sparse expert hybrid architecture and the Ministral 3 series (3B, 8B, 14B) for local and edge scenarios. According to the official introduction, Large 3 adopts a MoE structure of about 41B activation parameters and 675B total parameters, reaching the leading level of the current open weight model in general instruction tasks, multilingual dialogue and image understanding.
The Ministral 3 series also provides base, instruct, and reasoning variants, supporting multimodal and multilingual, focusing on outputting fewer tokens for the same task and achieving a better performance-cost ratio. Among them, the 14B reasoning version achieved leading results among similar-magnitude models on reasoning benchmarks such as AIME 2025. Mistral also collaborates with ecosystems such as NVIDIA, vLLM, and Red Hat to provide inference optimizations such as TensorRT-LLM and SGLang for the Mistral 3 family, as well as efficient deployment paths on Blackwell, Hopper GPUs, RTX PCs, and Jetson devices, so that the model can be extended from the data center to robots and edge terminals.
Mistral 3 is currently available on platforms such as Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM watsonx, OpenRouter, Fireworks, Unsloth AI, Together AI, and more, with plans to launch on NVIDIA NIM and AWS SageMaker。 Mistral also offers custom training services to build purpose-built models and enterprise-level deployments for organizations with domain-specific needs.
FAQs
Q: What is the Mistral 3 model family?
A: Mistral 3 is a new generation model series released by Mistral AI, including the large-scale Mistral Large 3 and Ministral 3 (3B, 8B, 14B) for local and edge scenarios.
Q: What are the technical features of the Mistral Large 3?
A: Large 3 adopts a sparse MoE architecture with about 41B active parameters and 675B total parameters, supporting image understanding and multilingual dialogue, and is at the leading level among open-source instruction models.
Q: What are the main applications for the Ministral 3 series?
A: Ministral 3 focuses on local and edge deployments, providing base, instruct, and reasoning variants to reduce token generation while ensuring effectiveness, making it suitable for inference and multimodal applications in resource-constrained environments.
Q: How are these models licensed?
A: The basic and directive versions of Mistral Large 3 and Ministral 3 are open-source under the Apache 2.0 license, making it easy for enterprises to fine-tune, deploy, and commercialize under the premise of compliance.
Q: Where can Mistral 3 be used and deployed today?
A: In addition to providing APIs in its own Mistral AI Studio, Mistral 3 has been integrated into multiple clouds and development platforms such as Amazon Bedrock, Azure Foundry, and Hugging Face, and plans to further expand to NVIDIA NIM and AWS SageMaker.
