On March 26th, the Chinese Institute of Electronics (CIE) announced the winners of its scientific and technological awards. The "Key Technologies and Applications of the Angel Machine Learning Platform for Large-Scale Data" jointly developed by Tencent, Peking University, and Beijing University of Science and Technology won the first prize for technological progress.
The Angel machine learning platform, as introduced, serves as the core technology supporting Tencent's Mixnet large-scale model. Its main objective is to address the challenges of training large models with massive amounts of data and difficult architecture design. It possesses industry-leading hardware acceleration and online inference service capabilities, making it an indispensable foundational platform for artificial intelligence model training.
The adjudication committee recognized that Tencent's Angel machine learning platform has high technical complexity, large research and development difficulty, strong innovation, and broad application prospects. Its overall technology has reached an international advanced level. Particularly, the efficient cache scheduling and management technology for all-to-all communication and the adaptive presampling and graph structure search technology have reached an internationally leading level.
Tencent's Angel machine learning platform supports the training of Tencent's Mixnet large model. Since its debut in September 2023, Tencent Mixnet has expanded its model to a trillion-level parameter scale by adopting a mixture of expert (MoE) structure, achieving performance improvement in prediction and cost reduction in inference. As a universal model, Tencent Mixnet leads the industry in Chinese performance, especially in mathematical deduction, logical reasoning, and multi-turn dialogue. Currently, Tencent Mixnet is also actively developing multimodal models to further enhance its abilities in text-to-image and text-to-video scenarios.
