Tencent recently unveiled its latest large language model, Hunyuan-T1, claiming its reasoning capabilities rival those of OpenAI's best reasoning systems.
According to Tencent, Hunyuan-T1's development heavily relied on reinforcement learning, with a remarkable 96.7% of post-training computing power dedicated to enhancing its logical reasoning and aligning it with human preferences.
Hunyuan-T1 demonstrated strong performance across various benchmark tests. In the MMLU-PRO evaluation, assessing knowledge across 14 academic disciplines, it achieved a score of 87.2, slightly below OpenAI's o1 model. In scientific reasoning, it scored 69.3 on the GPQA-diamond test.
Notably, Tencent highlighted Hunyuan-T1's exceptional mathematical abilities, achieving a score of 96.2 on the MATH-500 benchmark, second only to Deepseek-R1. Furthermore, the model excelled in code generation (LiveCodeBench: 64.9) and high-difficulty reasoning (ArenaHard: 91.9). Tencent also stated that Hunyuan-T1 boasts over 90% accuracy on numerous Chinese language tasks.
In model training, Tencent employed curriculum learning, gradually increasing task difficulty. They also innovatively developed a self-reward system, using earlier model versions to evaluate newer outputs, thus continuously driving performance improvements.
Hunyuan-T1 utilizes a Transformer Mamba hybrid architecture, which Tencent claims processes long texts twice as fast as traditional models under the same conditions. Currently, Hunyuan-T1 is available through Tencent Cloud and offers a demo on Hugging Face.
This launch marks another significant move in the AI competition among Chinese tech giants, following similar announcements from Baidu and Alibaba regarding their self-developed models purportedly reaching o1 level. It's noteworthy that Alibaba, Baidu, and Deepseek are actively promoting open-source strategies. AI investor and former Google China president, Kai-Fu Lee, previously stated that these Chinese AI models pose a potential existential threat to OpenAI.
