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

Chinese tech giant Tencent Holdings has rolled out its latest flagship artificial intelligence model, Hy4, seeking to challenge top open-weight rivals with enhanced agentic capabilities built to power its enterprise software ecosystem.

Released on Wednesday, the large language model (LLM) introduced visual understanding for the first time. Hy4 could process images, analyse complex charts and conduct visual mathematical reasoning, according to the Shenzhen-based company.

Tencent said the release marked the first time its AI line-up became “one of the strongest open-weight models available today”, although it acknowledged that the model still trailed the world’s leading closed-source systems, particularly on challenging internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering.

“We will continue to steadily scale pre-training and post-training to narrow this gap,” Tencent said in a blog post on Wednesday,

The company has been increasing both model size and training data to close the gap with competitors. The Hy4 series features 770 billion parameters – a 2.6-fold jump from the 295 billion parameters in its predecessor, Hy3.

Tencent indicated that its next-generation model, Hy5, would scale further, while a soon-to-come, lightweight Hy4 Flash version would slash task costs.

The company is also laying the foundation for recursive self-improvement in Hy4, expanding its ability to carry out autonomous AI research and development.

Underlining a “co-design” strategy to develop the AI model alongside products used in real enterprise workflows, Tencent said Hy4 powered a comprehensive update to its workplace AI tool WorkBuddy. The model achieved an average score of 83.18 and a pass rate of 96.15 per cent in internal evaluations for WorkBuddy tasks, according to the company.

The launch underscores Tencent’s efforts to hold its ground in China’s fierce AI race, with domestic rivals including DeepSeek, Z.ai and Alibaba Group Holding already challenging global peers in the open-weight landscape. Alibaba owns the South China Morning Post.

A preview version of Hy4, launched in late August, ranked 23rd globally on Wednesday on Code Arena’s WebDev leaderboard – a real-time benchmark evaluating LLMs on real-world coding tasks.

With significantly expanded training data and model scale compared with the preview release, Hy4 improved on its earlier version across 12 tasks spanning agentic capabilities, coding and workplace automation, according to Tencent.

The new flagship outperformed Z.ai’s GLM-5.3 and the DeepSeek-V4.1-Flash-1017 model across all tested tasks, while trading wins with Moonshot’s Kimi K3.1 Pro, according to benchmarks published in the model’s technical blog.

The Hy4 release also marks another step towards Tencent’s goal of building frontier, state-of-the-art AI models. The company aggressively increased its capital expenditure by 176 per cent year on year to 52.8 billion yuan (US$7.9 billion) in the second quarter to fund its AI efforts, pushing free cash flow into negative territory for the period.

Tencent also moved to mitigate cybersecurity concerns by restricting the hacking capabilities of Hy4 during training by filtering out content related to exploit development, reverse engineering and penetration testing “to avoid amplifying potential offensive capabilities”.