Hunyuan-T1

What it is

Hunyuan-T1 is Tencent's deep-reasoning model. It is built on the TurboS base and is a Hybrid-Transformer-Mamba mixture-of-experts model. Tencent called it the industry's first ultra-large-scale model of this kind ^[1][2].

Deep-reasoning models are trained to spend more computation on multi-step problems, such as mathematics and programming, before they answer. Reinforcement learning, a method that rewards outputs meeting a goal, is the usual way such reasoning is strengthened.

Origins

Its predecessor, T1-Preview, launched in February 2025 in the Tencent Yuanbao app ^[1]. Hunyuan-T1 was officially released on 21 March 2025 and made available on the Tencent Cloud website ^[3].

Architecture and training

Tencent said the Mamba design improves long-sequence processing and doubles decoding speed under the same deployment conditions ^[1]. Of the post-training compute, 96.7% went to reinforcement learning. Curriculum learning, data replay and periodic policy resets were used in training, and Tencent said these improved long-term training stability by more than 50% ^[1]. Reinforcement learning trains a model by rewarding outputs that meet a goal, and curriculum learning orders the training tasks from easier to harder.

Alignment combined self-rewarding, in which scoring was done by an early version of T1-Preview, with a reward model ^[1].

Reported scores

Tencent reported MMLU-Pro 87.2, GPQA-Diamond 69.3, LiveCodeBench 64.9, MATH-500 96.2 and ArenaHard 91.9 ^[1]. Tencent said the model roughly matches or slightly beats DeepSeek-R1 on several benchmarks ^[1].

China Internet Watch reported that it was priced at one-quarter of DeepSeek-R1's standard-time rate ^[2]. The launch API price was 1 yuan per million input tokens and 4 yuan per million output tokens, and Tencent cited decoding speeds of 60 to 80 tokens per second ^[3].

Access

Hunyuan-T1 was offered through a web demo and the Tencent Cloud API, and Tencent did not release its weights ^[1].

Place in the field

Hunyuan-T1 extended the Hunyuan-TurboS base into a reasoning model. Its predecessor, T1-Preview, had already been offered in the Yuanbao app before the official release ^[1].