The French company Mistral AI presented a new lineup of AI models Mistral 3, aimed at developers, enterprises, and the open-source community. The family includes compact models with 3, 8, and 14 billion parameters, as well as the flagship Mistral Large 3, which features a mixture of experts architecture with 41 billion active parameters. All models are available under the Apache 2.0 license and are already featured on platforms such as Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI.
Mistral Large 3 was trained on 3,000 NVIDIA H200 GPUs using Blackwell attention and mixture-of-experts kernels technologies. This enabled high performance in multilingual and multimodal tasks. Mistral 3 models offer base, instructive, and logical variants, each supporting image recognition and a balance of cost and performance.
The models are available in NVFP4 format, optimized for vLLM and NVIDIA hardware, ensuring efficient operation both in data centers and on edge devices. With open weights and broad language support, these models easily adapt to various workflows and are suitable for custom configurations in companies.
Mistral AI actively collaborates with NVIDIA, Red Hat, and vLLM to enhance the training and implementation of its solutions. Industry representatives highlight the code openness, multilingualism, and flexibility of the models as key advantages of the new lineup for developers and corporate users.
