Tencent Unveils Hunyuan Turbo S: A Leap in AI Speed and Efficiency, Challenging DeepSeek

Tech giant Tencent has launched its latest AI model, Hunyuan Turbo S, boasting faster query responses than the DeepSeek-R1 model. Available via API on Tencent Cloud, this model aims to set a new standard for AI efficiency.

Tencent announced on WeChat that Hunyuan Turbo S doubles the output speed and cuts down the initial response delay by 44%. They liken its rapid processing to human intuition, where quick decisions often trump slower, rational thought. However, the company clarifies that this new model solves problems effectively by integrating both quick and more thorough thinking processes. It uses a hybrid architecture, blending Mamba and Transformer technologies. This design enhances efficiency by reducing computational complexity, minimizing storage usage, and lowering training and operational costs. The model uses Mamba’s capability to handle long sequences while also utilizing Transformer’s ability to understand complex relationships within data.

Tencent claims this marks the first successful application of the Mamba architecture in a large Mixture of Experts (MoE) model without compromising performance. Benchmark tests released by Tencent suggest that Hunyuan Turbo S performs as well as, if not better than, models like DeepSeek-V3, Claude 3.5 Sonnet, and GPT-4o in tasks involving mathematics, coding, and reasoning.

The input API is priced at 0.8 yuan ($0.11) per million tokens, with output priced at 2 yuan ($0.28) per million tokens.

With DeepSeek’s rise, the competition in China’s AI sector is intensifying. Alibaba recently previewed its Qwen QwQ-Max reasoning model and committed $52 billion to AI infrastructure over the next three years. DeepSeek is also planning to release its next reasoning model, DeepSeek R2, possibly sooner than initially planned, with expectations of improved coding abilities and multilingual reasoning.