Mistral has launched Mistral 3, a new generation of AI models designed to be open, multimodal, and multilingual. This lineup includes three smaller models (14B, 8B, and 3B parameters) and the flagship Mistral Large 3, which features a mixture-of-experts architecture with 41B active and 675B total parameters. All models are released under the Apache 2.0 license, promoting accessibility for developers and enterprises. Mistral Large 3 is notable for its performance, achieving results comparable to top instruction-tuned models while excelling in image understanding and multilingual tasks. The models were trained using 3,000 NVIDIA H200 GPUs, marking a significant advancement in Mistral's pretraining techniques. They are optimized for deployment on various NVIDIA systems, enhancing their efficiency and performance. The collaboration with NVIDIA, vLLM, and Red Hat has resulted in a checkpoint format that allows efficient use of Mistral Large 3 on high-performance hardware. For edge applications, the Ministral 3 models come in three sizes and include base, instruct, and reasoning variants, with a strong focus on cost-to-performance ratios. Mistral 3 is now available on platforms like Mistral AI Studio, Amazon Bedrock, and Hugging Face, among others. For organizations needing customized solutions, Mistral AI offers services to fine-tune models for specific tasks. The article emphasizes the importance of transparency and collective progress in AI development, inviting users and developers to leverage these new models for diverse applications.