Tencent officially releases the Hunyuan-T1 reasoning model

Background

A reasoning model is a language model trained to work through a problem in a series of intermediate steps before it gives a final answer. Tencent's Hunyuan team had already tried this approach in February 2025, when the Hunyuan-T1-Preview model launched inside the Yuanbao app.^[2] The full Hunyuan-T1 model was built on that preview.^[2]

What happened

On 21 March 2025 Tencent officially launched the Hunyuan-T1 deep-reasoning model on Tencent Cloud.^[3] The model is built on Tencent's TurboS base model and uses a hybrid Transformer-Mamba architecture with a mixture-of-experts (MoE) structure.^[1] In a mixture-of-experts model the network is divided into many specialist sub-networks, and only some of them are active when the model processes each token. A model of this kind can hold a large total number of parameters while doing less work on each step. Mamba is a sequence-modelling architecture that handles long inputs differently from the attention mechanism used in standard Transformers.

Tencent positioned Hunyuan-T1 against DeepSeek-R1 and set its price at one-quarter of R1's standard-time rate.^[1] The release came shortly after Baidu's ERNIE X1.^[1]

Key details

The launch API price was 1 yuan per million input tokens and 4 yuan per million output tokens.^[3] Tencent reported 87.2 on MMLU-Pro, 69.3 on GPQA-Diamond and 96.2 on MATH-500.^[2] MMLU-Pro is a test of knowledge and reasoning across many subjects. GPQA-Diamond is a set of difficult graduate-level science questions. MATH-500 is a mathematics benchmark. These are the figures Tencent published for the model. Benchmark scores measure performance on fixed sets of questions, so they show how the model did on those tests rather than how it will perform on every task.

Why it matters

The release put a reasoning model on Tencent Cloud, where developers could call it through an API with published prices. Before this, the Hunyuan-T1-Preview had reached users through the Yuanbao app.^[2] Tencent's comparison with DeepSeek-R1 centred on price, which made the cost of using a reasoning model part of the product's pitch.^[1] For a general reader, the practical point is that a reasoning model works through more steps before it answers, and each step uses tokens, so the price per token matters when comparing models for regular use.