Just a month after having introduced his model of turbos reasoning, the Chinese conglomerate tencent reveals the one for whom he served as a basis: Hunyuan-T1. According to him, thanks to a large-scale post-training, his reasoning capacity has been considerably extended and aligned with human preferences, which allows him to compete with Deepseek R1.

From 2024, with V2, a model of efficient language offered at a competitive cost, Deepseek sparked a price war on the Chinese AI market and brought Tencent and its main competitors including Zhipu Ai, Bytedance, Alibaba, Baidu, to review their prices down. While the technological war around AI between the United States and China continues to intensify since the appearance of R1, competition in the Empire of the environment also reaches new heights.

A model focused on deep reasoning

After Baidu and Alibaba, it is the giant Tencent who tries to win on the Chinese market against Deepseek.

T1 is based on Hybrid-Transformer-Mamba Moe architecture, which as its name suggests, combines the advantages of MAMBA transformers and models, while integrating experts, which limits the number of active parameters. It is particularly suitable for tasks requiring long context treatment and great precision. T1 thus reduces context losses and optimizes the use of IT resources, while being twice as fast in decoding.

Thanks to a post-workout based on the RLHF (learning by strengthening with human return), Tencent positions its model as a serious competitor in front of Openai O1 and Deepseek R1.

According to the evaluations shared by Tencent, Hunyuan-T1 displays performance:

• Superior or equivalent on certain benchmarks (Mmlu-Pro, Ceval, Aime, Zebra Logic);

• Particularly strong in mathematics, with an impressive score of 96.2 on Math-500;

• Solid in engineering and coding, demonstrating an advanced capacity to solve technical problems.

Benchmarks provided by Tencent