Tencent Hunyuan Large⁣ Model Ranks First in China in SuperCLUE Evaluation Report

On September 2, SuperCLUE, a Chinese large model evaluation benchmark, ⁢released the “Chinese Large Model Benchmark Evaluation August‌ 2024 Report”. Tencent’s Hunyuan Large Model ranked first among domestic large models in⁢ total score due⁢ to its outstanding performance​ in multiple core ​tasks, ‌becoming one⁣ of the fastest-improving⁤ models on the list.

According to the SuperCLUE report, in the evaluation of 11 capabilities, ⁢Tencent Hunyuan ranked ‍first in China in 8 core tasks. “Tencent Hunyuan ⁤has good overall capabilities and‌ is a very competitive general-purpose‌ large model.”

What SuperCLUE is evaluating this time is Tencent Hunyuan’s new​ generation ⁢large language model preview version (Turbo-Preview). The model ⁤adopts a new hybrid expert model (MoE) structure, from training data, model architecture, training strategy, training framework, software and ​hardware system In other aspects, it has realized full-link self-research. On the one hand, the ‌model has achieved a significant improvement in performance, ⁣and on ‍the other hand, it has also achieved a significant ⁣reduction in reasoning costs,⁢ which has great application​ potential.

As an independent third-party Chinese large model benchmark evaluation organization, ‍SuperCLUE’s August report focuses on general ability evaluation, and the evaluation plan consists of three dimensions: science, liberal arts, and hard. Specifically, science abilities include calculation, logical reasoning, and coding ability; liberal arts tasks cover seven dimensions: knowledge‌ encyclopedia, language⁣ comprehension, long text, role-playing, generation‍ and creation, security, and tool use; and hard tasks⁣ focus on precise instruction following and high-level reasoning of complex tasks.

As the best model ‍in China, Tencent Hunyuan ranked first in both science and liberal arts. Tencent ​Hunyuan performed well in the Hard task, scoring 74.33 points, the only⁤ large model in China with⁢ a score of more than 70 points, only slightly behind ChatGPT-4o.

It is worth noting that with⁣ the vigorous development of the large‍ model industry, domestic large models represented⁣ by Tencent Hunyuan are accelerating​ their evolution and upgrading their ⁢capabilities. The ​evaluation report data ‍shows ‌that in general, the gap between the general capabilities of the top domestic large model in the Chinese​ field ‌and the leading foreign models continues to‍ narrow, from 30.12% in May 2023 to 1.29% in August 2024, with a small gap of‍ only about 1 point in the⁤ total score.

Since its official debut in September 2023, Tencent Hunyuan has been the first in China to adopt the MoE structure, and‍ the model has expanded to ​a trillion-parameter scale. The overall performance has been continuously⁣ upgraded. In addition ⁣to general capabilities and text-to-text, it also has outstanding performance in multimodal capabilities such ⁢as text-to-image, image-to-text,​ and video ⁢generation.‍ In the previously​ released Chinese multimodal large model SuperCLUE-V benchmark list, Tencent Hunyuan’s large model⁤ ranked first among domestic large models due to its ⁤outstanding performance in multimodal understanding, and has remained in the⁤ excellent⁣ leader quadrant.

Based on the accumulation‌ of leading model capabilities, ​Tencent Hunyuan Big Model is actively promoting the implementation ‌of applications to create more⁤ value for the big model. Currently, nearly 700 businesses and scenarios​ within Tencent have been connected, including Tencent Yuanbao, Tencent⁢ Cloud, QQ, WeChat Reading, Tencent News, Tencent Customer Service, etc. ⁤Previously, Tencent’s collaborative SaaS (software as a service) products were fully ​connected to the Tencent Hunyuan Big Model.

Tencent Hunyuan large model provides model services of various sizes⁤ on Tencent Cloud, and is fully open to enterprises and individual developers through⁤ access and use methods such as API, exclusive models, and fine-tuning models. Currently, the cloud​ versions ⁤of Tencent Hunyuan⁤ include Turbo-Preview, Pro, Standard, Lite and other versions; code generation, role-playing, Functioncall, etc. are‌ open on the exclusive‍ model; enterprises can also fine-tune Tencent Hunyuan through the Tencent Cloud TI ‍platform.

Based on years of experience and accumulation in industrial Internet, Tencent Cloud has joined hands with⁣ leading companies⁣ in the industry to output more than 50 solutions for more than 20 ⁣industries, providing a complete set of model service tool chains to​ help companies create and deploy AI applications efficiently, with high quality and low ‍cost.