The latest Chinese large model evaluation is released, and Tencent Hunyuan ranks in the Excellent Leaders Quadrant

Reporters reported on May 6 that SuperCLUE, an authoritative domestic large model evaluation organization, recently released the "Chinese Large Model Benchmark Test 2024 Fourth Quarter Report". Among them, Tencent Hunyuan large model ranks in the first echelon of domestic large models. It is in a leading position in both basic and scenario applications, and is located in the Excellent Leaders Quadrant.

#SuperCLUE is the domestic authoritative comprehensive evaluation benchmark for general large models. It was formerly the well-known third-party Chinese language understanding test benchmark CLUE (The Chinese Language Understanding Evaluation). SuperCLUE builds a multi-level, multi-dimensional comprehensive evaluation benchmark based on the wide application of general large models in academia, industry and users. It consists of ten basic tasks, including logical reasoning, coding, language understanding, long text, and role playing. wait.

This report selects the April versions of 32 representative large models at home and abroad. Through multi-dimensional comprehensive evaluation, it truly and accurately reflects the comprehensive capabilities and development of large models at home and abroad in the Chinese field. status quo. In terms of the total score ranking of the evaluation report, Tencent Hunyuan large model ranks among the top three, reflecting the strength of the leading model.

Among the top ten ability scores, Tencent Hunyuan large model has a relatively balanced ability. In terms of semantic understanding ability, it ranks first in the country with a score of 75.4; in role play, security Ability, calculation, logical reasoning, tool use, and long text ability are also among the best in the country.

View the full text. The first echelon of domestic large models has reached or is close to the international first-class level, including Tencent Hunyuan, Wen Xin Yi Yan, and Tong Shi Qian Wen. There are large models from major manufacturers, as well as representatives from large model start-ups such as GLM-4, Baichuan3, Moonshot and Minimax.

After in-depth understanding, the Hunyuan large model is a practical large model built by Tencent based on full-link independent controllable technology. Since its debut in September 2023, it has passed Through continuous iteration and practice, we have accumulated complete independent technologies from underlying computing power to machine learning platform to upper-layer applications.

Tencent has its self-developed Xingmai high-performance computing network, which can bring 10 times communication performance improvement to large AI models; in terms of training and inference framework, Tencent’s self-developed machine learning platform training speed is the mainstream framework 2.6 times, the cost of large model inference is reduced by 70% compared with the industry's mainstream framework; algorithmically, Tencent's Hunyuan large model is the first to adopt the hybrid expert model (MoE) structure, and the overall effect of the model is 50% higher than that of the previous generation model.

Recently, the research "Key Technologies and Applications of Angel Machine Learning Platform for Large-Scale Data" jointly completed by Tencent, Peking University and University of Science and Technology Beijing also won the first prize of the 2023 China Electronics Society Science and Technology Award, reflecting It reflects Tencent’s profound accumulation of self-developed technology.

In terms of application, Tencent's Hunyuan large model has supported the access of more than 400 businesses and scenarios within Tencent. Tencent's collaborative SaaS products are fully integrated into Hunyuan and have achieved intelligent upgrades. Tencent Hunyuan has also been fully opened to enterprises and individual developers through Tencent Cloud.

Currently, the parameters of Tencent’s Hunyuan large model exceed one trillion, and the number of tokens exceeds 7 trillion. Previously, the "2024 China Large Model Capability Evaluation" released by Sullivan, an authoritative international research organization, showed that Tencent Hunyuan has ranked first in the domestic echelon in terms of general basic capabilities and professional application capabilities.

The above is the detailed content of The latest Chinese large model evaluation is released, and Tencent Hunyuan ranks in the Excellent Leaders Quadrant. For more information, please follow other related articles on the PHP Chinese website!