华为云发布盘古大模型5.5 新一代昇腾AI云服务上线

Core Insights - Huawei has officially launched the Pangu Model 5.5, which includes comprehensive upgrades to five foundational models: Natural Language Processing (NLP), Computer Vision (CV), Multimodal, Prediction, and Scientific Computing [1][2][3][4][5] - The Pangu models have been implemented across over 30 industries and 500 scenarios, demonstrating significant value in sectors such as government, finance, manufacturing, healthcare, mining, steel, railways, autonomous driving, and meteorology [1] Natural Language Processing (NLP) Model - The new 718 billion parameter deep thinking model is composed of 256 experts and significantly enhances capabilities in knowledge reasoning, tool invocation, and mathematics [1] - The model features upgrades in efficient long sequences, reduced hallucination, and the integration of fast and slow thinking, improving user experience and increasing reasoning efficiency by 8 times [1] Prediction Model - The Pangu Prediction Model utilizes a pioneering triplet transformer architecture to unify and efficiently process data from various industries, enhancing prediction accuracy and cross-industry generalization [2] Scientific Computing Model - The Pangu Scientific Computing Model has been integrated with various scientific applications, such as weather forecasting, improving accuracy and reducing errors in predictions [2] Computer Vision (CV) Model - Huawei has released a new 30 billion parameter CV model, the largest in the industry, supporting diverse data types for perception, analysis, and decision-making [3] - The model enhances the identification and precision of business scenarios by creating a library of rare visual fault samples across various industrial contexts [3] Multimodal Model - The new multimodal model can generate training data for intelligent driving and embodied intelligent robots by creating digital physical spaces, significantly reducing the need for costly road data collection [3] AI Cloud Service - The new generation of Ascend AI Cloud Service, based on CloudMatrix 384 super nodes, significantly enhances computing power, achieving a throughput of 2300 Tokens/s, nearly quadrupling performance compared to non-super nodes [4] - The super node architecture supports parallel reasoning for mixed expert models and optimizes resource allocation, improving effective utilization of computing power by over 50% [4] Scalability and Client Adoption - The AI cloud service can scale to support training tasks with trillions of parameters, linking up to 432 super nodes into a massive cluster of up to 160,000 cards [5] - Over 1,300 clients, including iFlytek, Sina, and the Chinese Academy of Sciences, are leveraging the Ascend AI Cloud Service to accelerate the intelligent upgrade across various industries [5]