Chinese technology giant Tencent has released its new AI model, Hunyuan Turbo S, which it claims can answer queries faster than the DeepSeek-R1 model. The model is available on the official Tencent Cloud website and is accessible via API.
Hunyuan Turbo S doubles the output speed and reduces the first word delay by 44 per cent, the company announced on its official WeChat channel.
Tencent said that the rapid thinking model is analogous to human intuition, which often results in quick responses compared to rational thinking. However, the company said that Hunyuan Turbo S efficiently solves problems by merging long and short chains of thought.
The model uses an innovative hybrid-mamba transformer fusion architecture. It optimises efficiency by reducing the computational complexity of the conventional transformer, minimising the use of KV-Cache storage and reducing training and inference costs.
The company also stated that the model leverages Mamba’s efficiency in processing long sequences, while preserving Transformer’s power in capturing complex contextual relationships.
Tencent also claims that this is the first time the Mamba architecture has been applied losslessly to a super-large Mixture of Experts (MoE) model.
Tencent also released benchmark results and the model is better, if not on par with other large language models such as DeepSeek-V3, Claude 3.5 Sonnet and GPT-4o, in mathematics, coding and reasoning tasks.
The input API price of Hunyuan Turbo S is 0.8 yuan ($0.11) per million tokens and its output price is 2 yuan ($0.28) per million tokens.
With the rise of DeepSeek, competition in the Chinese AI ecosystem is heating up. Recently, Alibaba unveiled a preview of the Qwen QwQ-Max reasoning model and pledged to invest $52 billion in AI infrastructure over the next three years.
It was also reported that DeepSeek plans to release its next reasoning model, DeepSeek R2, ‘as soon as possible’. The company had initially planned to release it in early May, but is now considering an earlier timeline.
The model is expected to produce ‘better coding’ and reasoning in languages other than English.
Translated with DeepL.com (free version)
Leave a Reply