Tencent releases Hunyuan Turbo S, a fast-thinking model

Background

Tencent positioned Hunyuan Turbo S as a fast-thinking model, and compared its reply speed directly with DeepSeek-R1, a reasoning model from DeepSeek ^[1].

Reasoning models spend extra computation working through intermediate steps before they answer, which usually makes them slower and more expensive than models that answer directly. A fast-thinking model aims to narrow that gap.

What happened

On 27 February 2025 Tencent released Hunyuan Turbo S, saying that it replies faster than DeepSeek-R1 ^[1]. Tencent said the model doubles word speed and cuts first-word delay by 44% ^[2]. Access went through a Tencent Cloud API with a waiting list and a free one-week trial, and the weights were not released ^[2].

Key details

The model was described as combining Mamba and Transformer layers in a super-large mixture-of-experts model ^[2]. Its API pricing was 0.8 yuan per million input tokens and 2 yuan per million output tokens ^[2]. Tencent reported an MMLU score of 89.5 and a Chinese-SimpleQA score of 70.8 ^[2].

Chinese-SimpleQA is a set of short factual questions in Chinese, so the score reflects factual recall in that language.

Impact

Because access was through a cloud API and the weights stayed unreleased, the launch put Hunyuan Turbo S in front of developers as a paid, hosted service rather than as a model that anyone could download. For developers, the access terms and the per-token price were the most immediate details. For researchers, the performance figures come from Tencent's own report and are best checked against independent evaluations.

The access terms set the practical pace of testing: a waiting list and trial period affect how quickly a team can evaluate the model, and the per-token price determines the cost of running it at scale.

For a reader, the main point is that the model was offered as a service rather than as an open download, which shapes how it can be used.