Tencent unveils Hy4 as open-weight AI race with DeepSeek, Z.ai, Alibaba heats up

Chinese technology firm Tencent has introduced its newest AI model, Hy4, aimed at competing with leading open-weight models like DeepSeek, Z.ai, and Alibaba Group Holding. Hy4 boasts advanced agentic abilities, enabling it to power Tencent's enterprise software ecosystem. A major update to the model is its first-time visual understanding, allowing it to process images, analyze complex charts, and perform visual mathematical reasoning.

While Tencent claims Hy4 is now among the top-performing open-weight models, it acknowledges that the AI still lags behind some closed-source systems, especially on tasks related to research-level mathematics, long-horizon coding, model self-improvement, and visual question answering. The company plans to bridge this gap by steadily increasing pre-training and post-training to bolster Hy4's performance.

Hy4 has 770 billion parameters, a nearly threefold increase from the 295 billion in its predecessor, Hy3. The company aims to continue scaling Hy4 larger and adding more training data to narrow the gap with competitors. A future model, Hy5, is reportedly in the works, and a lightweight Hy4 Flash version will be released to further reduce task costs. Moreover, Tencent is working on recursive self-improvement to allow Hy4 to carry out autonomous AI research and development.

Hy4 is integrated into Tencent's WorkBuddy, an AI-powered workplace tool, and has shown impressive results in internal evaluations. The model scored an average of 83.18 and achieved a 96.15% pass rate in WorkBuddy tasks. These accomplishments highlight Tencent's efforts to remain competitive in China's vigorous AI race, with domestic rivals DeepSeek, Z.ai, and Alibaba Group also vying for dominance in the open-weight AI landscape.

A preview version of Hy4, launched in late August, ranked 23rd globally on Code Arena's WebDev leaderboard, a real-time benchmark that assesses LLMs on real-world coding tasks. With improved training data and model size compared to the preview release, Hy4 outperformed Z.ai's GLM-5.3 and DeepSeek-V4.1-Flash-1017 in 12 tasks encompassing agentic capabilities, coding, and workplace automation.

The new model also trailed Moonshot's Kimi K3.1 Pro in some tests. The Hy4 launch represents another step towards Tencent's ambition of constructing frontier, state-of-the-art AI models. In pursuit of this goal, the company increased its capital expenditure by 176% year-on-year to 52.8 billion yuan (US$7.9 billion) in the second quarter, resulting in negative free cash flow for the period.

To address cybersecurity concerns, Tencent restricted Hy4's hacking capabilities during training to filter out content related to exploit development, reverse engineering, and penetration testing.