(Aug 28): Tencent Holdings Ltd released a foundation model it says outperforms rivals Z.AI Co and Moonshot AI in internal tests, claiming a place among China’s top artificial intelligence players.

The Hy4 Preview model runs on 770 billion parameters and features a one million-token context window, the company said in a statement, marking a big improvement in size and data scale from its predecessor. The larger the context window, the more an AI model remembers and can respond to a user’s queries and commands.

In blind engineering tests conducted by Tencent’s own experts, the Hy4 Preview beat Z.ai’s GLM-5.3 and Moonshot’s Kimi K3 by a narrow margin. On third-party benchmarks, Hy4 Preview’s scoring varied against domestic rivals including Alibaba Group Holding Ltd’s Qwen 3.8 Max. The company also showcased video demos using the model for scenarios such as game creation, 3D web design and accounting compliance workflows.

Tencent released the weights — the values that help AI systems make decisions — for its latest model. They can be accessed via its cloud platform and integrated into products such as its WorkBuddy tool and chatbot. A later edition of Hy4 will improve its capabilities through further pre- and post-training, Tencent said.

The Shenzhen-based social media giant — regarded as a laggard in the global race to build powerful AI models — is counting on its WeChat and gaming platforms to bankroll research and the increasingly expensive chips required for AI. The company hiked capital expenditures by 176% on AI projects and computing to 52.8 billion yuan (US$7.9 billion or RM31.66 billion) in the quarter ended June.

Its efforts are showing signs of traction through platforms like WorkBuddy — now China’s most popular desktop office agent — prompting a counteroffensive by ByteDance Ltd and Alibaba in the enterprise AI market.

Uploaded by Chng Shear Lane