Hunyuan Turbo S

What it is

Hunyuan Turbo S is a fast-thinking model that Tencent released on 27 February 2025. Tencent said it answers faster than DeepSeek-R1 ^[1]. It is described as a hybrid Mamba-Transformer design inside a very large mixture-of-experts model ^[2]. Mamba is a sequence-model architecture that works alongside attention layers, and combining it with Transformer layers is the design choice that sets this model apart.

Architecture and training

The Hunyuan-TurboS technical report, published in May 2025, describes a model with 56 billion activated and 560 billion total parameters. It has 128 layers that mix Mamba2, attention and MoE feed-forward layers ^[3].

The model was pre-trained on 16 trillion tokens with a 256K context. The report calls it the first industry-deployed large-scale Mamba model ^[3]. Post-training used 3 million instructions, adaptive fusion of long and short chain-of-thought reasoning, and two-stage large-scale reinforcement learning ^[3]. Chain-of-thought means the model works through intermediate reasoning steps before it gives an answer.

Main features

Tencent said the model doubles word-generation speed and cuts first-word latency by 44% compared with previous models ^[2]. Latency here is the wait before the first word of a reply appears.

At launch its API price was 0.8 yuan per million input tokens and 2 yuan per million output tokens, much cheaper than earlier Turbo models ^[2].

Reported launch scores were MMLU 89.5, slightly above GPT-4o, and Chinese-SimpleQA 70.8, against 68.0 for DeepSeek. The model lagged GPT-4o and Claude 3.5 on SimpleQA and LiveCodeBench ^[2]. The technical report cites an LMSYS Chatbot Arena score of 1356, which placed the model in the overall top seven, and a 77.9% average across 23 automated benchmarks ^[3].

Access

At launch it was offered through the Tencent Cloud API with a waiting list and a one-week free trial. It was not available for download ^[2].

Place in the field

Hunyuan-TurboS became the base of Tencent's Hunyuan-T1 reasoning model ^[3].