Introducing Mistral 3

Mistral AI has officially launched Mistral 3, an open model family that spans large frontier models and small edge models. The core of this release includes Mistral Large 3, as well as Ministral 3 in three sizes: 3B, 8B, and 14B. All models are released under the Apache 2.0 license, and for developers, enterprises, and the research community, the key point is not just that they can be tried out, but that they can be deployed, fine-tuned, and further developed under a relatively permissive license. Mistral Large 3 is currently Mistral's most powerful model, using a sparse mixture-of-experts architecture with 675B total parameters and 41B active parameters per inference, and it is offered in both base and instruction fine-tuned versions. The company states that it can compete with major open-weight instruction models on general instruction tasks, while also possessing image understanding and multilingual conversation capabilities, and that it ranks highly in the OSS non-reasoning category on LMArena. On the deployment side, Mistral emphasizes partnerships with NVIDIA, vLLM, Red Hat, and others, providing an NVFP4 checkpoint that supports more efficient execution on Blackwell NVL72, 8xA100, or 8xH100 nodes, and it mentions inference support such as TensorRT-LLM and SGLang. The smaller Ministral 3 series targets local and edge scenarios, with each size offering base, instruct, and reasoning versions, and featuring multimodal and multilingual capabilities; the company particularly emphasizes its cost-effectiveness, stating that the instruct models can achieve comparable or better performance with fewer output tokens in the same class, while the reasoning versions can trade longer reasoning for accuracy. In terms of availability, Mistral 3 is already available on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI, while NVIDIA NIM and AWS SageMaker are expected to support it later. For Taiwanese developers and researchers, the value of this article lies in the fact that Mistral 3 puts open weights, enterprise deployment, edge inference, and multimodal capabilities into the same product line, offering a complete range of choices from small on-device models to large MoE models.