[ad_1]
Tencent Holdings Ltd. unveiled an upgraded AI reasoning model, the latest bid by China’s most valuable company to compete for leadership in the country’s crowded artificial intelligence field.
The Shenzhen-based company rolled out the latest iteration of Hunyuan T1, which debuted last month in its chatbot app Yuanbao. T1 operates similarly to DeepSeek’s R1, the model that stunned Silicon Valley by matching the best from OpenAI and Meta Platforms Inc. at a fraction of the cost and computing power.
A team of Tencent engineers claimed in an event streamed online that the new model has cutting-edge logic and writing capabilities, with few hallucinations or distortions. T1’s performance is on par with state-of-art models including R1 and OpenAI’s o1, they said, citing industry benchmarks.
Tencent has joined Chinese tech peers including Alibaba Group Holding Ltd. and Baidu Inc. to piggyback on DeepSeek’s success to win new users. Baidu last week rolled out a reasoning model, which it says delivers performance on par with R1 at only half of the price, after having introduced DeepSeek’s model in its own search engine and maps.
Alibaba’s shares have surged 60% this year as the e-commerce pioneer pivoted toward AI, winning recognition for its open-source Qwen models. The company is also working with Apple Inc. to incorporate its AI technology to iPhones in China.
Tencent, the world’s biggest games publisher, has adopted a two-prolonged approach to generative AI. The company will rely on both in-house projects like Hunyuan and open-sourced models, in much the same way it publishes both its own and third-party games.
Tencent’s chatbot Yuanbao, for example, integrates the open-sourced R1, which at one point helped the service become the country’s most downloaded free app on the iPhone. R1 also powers a wide suite of Tencent’s consumer-facing applications, including its games and WeChat super-app.
On Tuesday, Tencent released five new AI models that turn text or images into 3D visuals — tools that could help game studios streamline the time it takes to get a title to market. Last month, it unveiled the Hunyuan Turbo S, designed to respond instantly and distinguish itself from the deep reasoning approach of DeepSeek’s eponymous chatbot.
During the live demo, Tencent tested T1 and R1 on the same complex math problems, which showed T1 returned results much more quickly than R1. Tencent said T1’s decoding speed is twice of those models with comparable parameters, thanks to the architecture called Hybrid-Mamba-Transformer.
Near the end of the session, Tencent said it will soon open-source one of its mid-scale models, whose performance is comparable to its flagship model.
Disclaimer: This story has been published from a wire agency feed without modifications to the text.
[ad_2]
Source link
