英伟达官宣新合作成就：Mistral开源模型提速，任意规模均提高效率和精度

Core Insights - Nvidia has announced a significant breakthrough in collaboration with French AI startup Mistral AI, achieving substantial improvements in performance, efficiency, and deployment flexibility through the use of Nvidia's latest chip technology [1] - The Mistral Large 3 model has achieved a tenfold performance increase compared to the previous H200 chip, translating to better user experience, lower response costs, and higher energy efficiency [1][2] - Mistral AI's new model family includes a large frontier model and nine smaller models, marking a new phase in open-source AI and bridging the gap between research breakthroughs and practical applications [1][6] Performance Breakthrough - Mistral Large 3 is a mixture of experts (MoE) model with 67.5 billion total parameters and 41 billion active parameters, featuring a context window of 256,000 tokens [2] - The model utilizes Wide Expert Parallelism, NVFP4 low-precision inference, and the Dynamo distributed inference framework to achieve best-in-class performance on Nvidia's GB200 NVL72 system [4] Model Compatibility and Deployment - The Mistral Large 3 model is compatible with major inference frameworks such as TensorRT-LLM, SGLang, and vLLM, allowing developers to deploy the model flexibly across various Nvidia GPUs [5] - The Ministral 3 series includes nine high-performance models optimized for edge devices, supporting visual functions and multi-language capabilities [6] Commercialization Efforts - Mistral AI is accelerating its commercialization efforts, having secured agreements with major companies, including HSBC, for model access in various applications [7] - The company has signed contracts worth hundreds of millions of dollars and is collaborating on projects in robotics and AI with organizations like the Singapore Ministry of Home Affairs and Stellantis [7] Accessibility of Models - Mistral Large 3 and Ministral-14B-Instruct are now available to developers through Nvidia's API directory and preview API, with all models accessible for download from Hugging Face [8]