Recently, Tencent officially launched its self-developed deep thinking model, the official version of Hunyuan T1.
Based on large-scale reinforcement learning, the official version of Hunyuan T1 has been specifically optimized for science and engineering challenges such as mathematics, logical reasoning, science, and code, resulting in a significant improvement in reasoning ability. On common benchmarks, such as the MMLU-PRO enhanced dataset for large language model evaluation, Hunyuan T1 achieved an excellent score of 87.2, second only to the top model o1. Simultaneously, in public benchmark tests like CEval, AIME, and Zebra Logic, which cover Chinese and English knowledge and competition-level mathematics and logical reasoning, Hunyuan T1 demonstrated the level of a leading reasoning model in the industry.
Beyond basic reasoning capabilities, the official version of Hunyuan T1 has also demonstrated strong adaptability in various alignment tasks, instruction-following tasks, and tool utilization tasks. This is attributed to its innovative architecture, inherited from Hunyuan Turbo S, and the adoption of the Hybrid-Mamba-Transformer fusion model. This marks the first time in the industry that a hybrid Mamba architecture has been seamlessly applied to an ultra-large reasoning model, effectively reducing the computational complexity of traditional Transformer structures, decreasing KV-Cache memory usage, and significantly lowering training and inference costs.
Furthermore, based on its outstanding long-text capturing ability, Hunyuan T1 effectively addresses common issues of context loss and long-distance information dependence in long-text reasoning. The Hybrid Mamba architecture is specifically optimized for long-sequence processing, achieving efficient computation to ensure the ability to capture information from long texts while significantly reducing resource consumption. With a similar number of activation parameters, Hunyuan T1 achieves a 2x increase in decoding speed.
Currently, Tencent Hunyuan T1 is open for experience and has launched API services. Users can enjoy the convenience and efficiency of this powerful reasoning model with an input price of 1 yuan per million tokens and an output price of 4 yuan per million tokens, depending on their needs.
