.png){kind=logo}
Tencent, a Chinese tech behemoth, has shown a new AI model and claims that it can answer questions more quickly than DeepSeek's highly regarded R1. This underscores the mounting pressure on big rivals as the firm achieves both domestic and international success.
Tencent's new AI model, Hunyuan Turbo S, is rolled out with rising competition spearheaded by the high growth in DeepSeek as the latter continued to put massive pressure on major Chinese technology players in both regional and global scenes.
In a statement to Reuters, Tencent asserted that Hunyuan Turbo S can answer questions in less than a second, making it distinct from DeepSeekR1, Hunyuan T1, and other models that “have to think for a while before they can answer.” The company also pointed out substantial deployment expenses for Turbo S. Tencent provided test results showing that Hunyuan Turbo S matched DeepSeek's V3 model in normal AI benchmarks.
Tencent claims that its latest AI model, Hunyuan Turbo S, has demonstrated performance comparable to top models on the market, including DeepSeek-V3, OpenAI’s GPT-4o, and Anthropic PBC’s Claude, particularly in benchmarks for math, reasoning, and knowledge.
The quick triumph of Beijing-based DeepSeek's AI models, R1 and V3, has been the talk of the town over the past few months, propelling AI developers to speed up the rollout of rival models. Alibaba Group Ltd., for instance, has joined the fray with models like Qwen 2.5-Max, which it asserts performs better than DeepSeek V3.
To create Hunyuan Turbo S, Tencent employed a Hybrid-Mamba-Transformer mixture, improving the computational efficiency of the model and minimising the KV-Cache of the Transformer model. The produced model combines Mamba deep-learning architecture, best suited for processing long sequences, with the capacity of the Transformer to understand sophisticated contextual relationships in data.
This is also the first time that the industry has been able to apply the Mamba architecture to an ultra-large Mixture of Experts (MoE) model without loss," Tencent said in a statement.
The Mixture of Experts is a machine learning method that divides AI models according to specialised knowledge, enabling them to work together on solving problems.
Tencent highlighted that the new architecture meaningfully reduces the cost of training and deployment. With other competitive models such as DeepSeek making high-performance models available at less cost, Tencent and other companies have been pressured to lower service costs and conduct more efficient research into AI training and inference algorithms.
As a flagship model, Tencent indicated that the Hunyuan Turbo S would be a basic foundation for eventual inference, text generation, and code generation upgrades.
