Tencent unveils new AI model 'Hunyuan T1' that rivals DeepSeek R1 in performance and price



Tencent, a major Chinese IT company, released the large-scale language model ' Hunyuan T1 ' on March 22, 2025. Tencent claims that Hunyuan T1 is a 'breakthrough in AI inference.'

Tencent Conflict

https://llm.hunyuan.tencent.com/#/blog/hy-t1




Tencent's Hunyuan T1 AI reasoning model rivals DeepSeek in performance and price | South China Morning Post
https://www.scmp.com/tech/big-tech/article/3303456/tencents-hunyuan-t1-ai-reasoning-model-rivals-deepseek-performance-and-price

On March 22, 2025, Tencent announced a new large-scale language model, 'Hunyuan-T1.' This model is based on 'Hunyuan TurboS,' which Tencent released in early March and specializes in high-speed processing, and can generate answers within one second.

The preview version of 'Hunyuan T1-Preview' released in mid-February was based on the medium-sized Hunyuan, so the 'Hunyuan-T1' based on TurboS has significantly improved overall performance. In particular, thanks to the long-text processing capability of ' Hunyuan TurboS ', Hunyuan-T1 can accurately grasp the context and process efficiently even when long text information is input. The official version of Hunyuan-T1 has significantly reduced the consumption of computing resources, and the decoding speed under the same conditions is twice as fast as the preview version.

When training Hunyuan T1, we collected a wide variety of problems from around the world, including mathematics, logical reasoning, science, and code, and used a curriculum learning approach to gradually increase the difficulty of the data and gradually expand the length of the model's context to improve its reasoning ability.

As a result, Hunyuan T1 not only performed comparable or slightly better than DeepSeek R1 on a variety of public benchmarks, including MMLU-pro, CEval, AIME, Zebra Logic, and other Chinese and English knowledge, competition-level math and logical reasoning metrics, but also performed comparable to DeepSeek R1 on human-evaluated datasets.

Below are the results of comparing Hunyuan T1, DeepSeek R1, GPT-4.5, and OpenAI o1 in each benchmark. Hunyuan T1 has the highest performance in the knowledge benchmark MMLU PRO and the inference benchmark DROP F1, and performs equivalent to DeepSeek R1 in CEval and CMMLU, which test Chinese knowledge.



Hunyuan T1 also offers excellent cost performance, with an input rate of 1 yuan (about 20 yen) per 1 million tokens and an output rate of 4 yuan (about 82 yen) per 1 million tokens. DeepSeek R1 also offers a similar input rate, with output rates of 16 yuan (about 330 yen) per 1 million tokens during the day and 4 yuan (about 82 yen) per 1 million tokens at night.

Tencent has released a demo of the Hunyuan T1. To use the demo, go to the following link:

Tencent Conflict

https://llm.hunyuan.tencent.com/#/chat/hy-t1



First, enter the prompt in the input field and click 'Send'. This time, I asked them to 'Write a short paragraph where the 1st letter of each sentence spells out the word 'CODE'. The message should appear natural and not obviously hide this pattern.'



The answer was generated in just a few seconds. Not only was it able to correctly generate sentences containing 'C', 'O', 'D' and 'E', but it also displayed the reasoning process in Chinese when answering.



On the other hand, when asked 'what happened during the Tiananmen Square Massacre,' Hunyuan T1 refused to answer, saying, 'Sorry, I haven't learned enough about how to answer this question yet, so I can't provide any information at this time.'



It has been noted that not only the Hunyuan T1 but also the DeepSeek R1 generate similar answers to questions dealing with such sensitive topics for the Chinese government.

'DeepSeek-R1' refuses to answer 85% of sensitive topics about China, but points out that restrictions can be easily circumvented - GIGAZINE

Related Posts:

in Review,   Software, Posted by log1r_ut