Tencent releases a new generation of large model "Hunyuan Turbo": reasoning efficiency increased by 100%, cost reduced by 50%

September 5,TencentAnnouncing the launch of a new generationLarge ModelHunyuan Turbo"Compared with the previous generation model, Tencent Hunyuan Turbo has significantly improved performance, with training efficiency increased by 108%, inference efficiency increased by 100%, inference cost reduced by 50%, and decoding speed increased by 20%. The effect is comparable to GPT-4o in multiple benchmarks, and ranks first in China in third-party evaluations.

Tencent releases a new generation of large model "Hunyuan Turbo": reasoning efficiency increased by 100%, cost reduced by 50%

At the Tencent Global Digital Ecosystem Conference, Qiu Yuepeng, Vice President of Tencent, COO of the Cloud and Smart Industries Group and President of Tencent Cloud, announced that Tencent Hunyuan Turbo has been launched on Tencent Cloud. The input and output prices are only half of the previous generation model, and enterprises and developers can directly access and use it on the cloud.

Currently, Tencent Hunyuan provides model services of various sizes on Tencent Cloud, which are fully open to enterprises and individual developers through access and use methods such as API, exclusive models, and fine-tuning models. Tencent Hunyuan provides multiple versions such as Turbo, Pro, Standard, and Lite on the cloud; 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.

Tencent releases a new generation of large model "Hunyuan Turbo": reasoning efficiency increased by 100%, cost reduced by 50%

(Figure: Public benchmark evaluation of Tencent Hunyuan Turbo and comparison with major domestic and foreign models)

Since last year, Tencent Hunyuan has been the first in China to adopt the MoE structure and has continued to upgrade on this technical route. Through its self-developed trillion-level inter-layer heterogeneous MoE structure, it uses different numbers of experts and different activation parameters in different layers of the model, while optimizing the training data, enabling the new generation of model Hunyuan Turbo to achieve significant improvements in both effect and performance.

In terms of industry-recognized benchmark indicators, Tencent Hunyuan Turbo is in the leading position in the domestic industry, and its performance is close to that of the top foreign models GPT4o and Claude3.5. As a new generation flagship large model, Tencent Hunyuan Turbo has made great improvements in language understanding, text creation, mathematics and code. Compared with the previous generation model, its complex mathematics solving ability has increased by 38%, and its code ability has increased by 32%.

Tencent releases a new generation of large model "Hunyuan Turbo": reasoning efficiency increased by 100%, cost reduced by 50%

(Figure: Public benchmark evaluation of Tencent Hunyuan Turbo and comparison with major domestic and foreign models)

On September 2, SuperCLUE, a benchmark for Chinese large models, released the "Chinese Large Model Benchmark Evaluation August 2024 Report". Tencent Hunyuan Turbo ranked first among domestic large models in total score thanks to its outstanding performance in multiple core tasks. As the best model in China, Tencent Hunyuan Turbo ranked first in both science and liberal arts. In Hard tasks centered around complex tasks and high-level reasoning, Tencent Hunyuan Turbo performed well, scoring 74.33 points, making it the only large model in China with a score of over 70, with only a slight gap from ChatGPT-4o.

As Tencent's self-developed large model, since its official debut in September 2023, Tencent Hunyuan has accumulated independent technologies from underlying computing power to machine learning platforms to upper-level applications through continuous iteration and practice. Its industry-leading technical strength has been recognized by many parties. In the selection of the 2023 Science and Technology Awards of the China Electronics Society, Tencent Hunyuan's "Key Technologies and Applications of Angel Machine Learning Platform for Large-Scale Data" won the first prize for scientific and technological progress.

Based on the leading accumulation of model capabilities, Tencent Hunyuan Big Model is actively promoting the implementation of internal 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.

statement:The content is collected from various media platforms such as public websites. If the included content infringes on your rights, please contact us by email and we will deal with it as soon as possible.
Information

The first end-to-end voice model in China, Lingo, was officially launched at the Bund Conference

2024-9-6 9:35:34

Information

Mianbi Intelligent released the MiniCPM 3.0 client-side model: it can run with 2GB of memory and its performance exceeds GPT-3.5

2024-9-6 9:38:33

Search