Tencent Releases Official Version of Self-Developed Deep Thinking Model Huiyuan T1

Tencent has recently launched the official version of its self-developed deep-thinking model, HunYuan T1. Known for its rapid response, quick output, and expertise in handling ultra-long documents, this model represents Tencent's powerful reasoning capabilities. Currently, HunYuan T1 is available on Tencent Cloud and is planned to be gradually rolled out on Tencent Yuanbao.

HunYuan T1’s official version leverages large-scale reinforcement learning and includes specialized optimizations for complex topics such as mathematics, logical reasoning, science, and coding. This further enhances its reasoning abilities. In various benchmark tests, HunYuan T1 has demonstrated exceptional performance. For example, it scored 87.2 points on the MMLU-PRO dataset, a widely recognized evaluation benchmark for large language models, just behind the leading models. Additionally, it achieved industry-leading results in public benchmarks like CEval, AIME, and Zebra Logic, which assess knowledge and competition-level mathematical and logical reasoning in both Chinese and English.

Furthermore, HunYuan T1 shows strong adaptability across multiple alignment tasks, instruction-following tasks, and tool-utilization tasks. This is due to its innovative architecture inherited from HunYuan Turbo S: the Hybrid-Mamba-Transformer fusion model. This structure marks the first industrial application of a hybrid Mamba architecture in a super-large reasoning model without compromising performance. Effectively, it reduces the computational complexity of traditional Transformer structures, minimizes KV-Cache memory usage, and significantly cuts down training and inference costs.

In the field of ultra-long text reasoning, HunYuan T1 exhibits distinct advantages. With its superior ability to capture long texts, it efficiently addresses common issues in long-text reasoning, such as context loss and long-distance information dependency. The hybrid Mamba architecture has been specifically optimized for long-sequence processing, enabling HunYuan T1 to double its decoding speed under similar activation parameters while drastically reducing resource consumption.

HunYuan T1 is now available on Tencent Cloud’s official website, offering API services. The input pricing is set at 1 yuan per million tokens, while the output cost is 4 yuan per million tokens.