Tencent Hy4 takes aim at leading open-source AI models with visual AI and computer-use features

Tencent has introduced Hy4, a new large language model that adds visual understanding and computer-use capabilities to its AI portfolio. The official post explains that the company plans to publish the model’s weights and system cards in the coming days, alongside safety measures focused on cybersecurity and biological risks.

Hy4 can interpret images, analyze complex charts, and tackle visual math problems, according to Tencent. It can also use command-line tools and graphical interfaces, allowing it to complete tasks inside software rather than only producing text. Tencent says it trained the model on professional applications including Blender and CAD software to improve planning for longer, multi-step assignments.

Focus on workplace tasks

The company is positioning Hy4 for office work, finance, software development, game creation, and scientific research. In Tencent’s WorkBuddy workplace assistant, the model can draw on information from chats, email, documents, calendars, and meetings. Tencent says this enables it to track projects, identify unfinished tasks, flag meetings requiring preparation, and monitor deadlines.

For financial work, Tencent says Hy4 performed best among the compared open-source models in a blind test involving 213 finance professionals and 579 practical tasks. The company reports that the model’s hallucination rate in internal tests fell to 7.2 percent, from 12.95 percent for Hy4 Preview. Such company-run evaluations provide useful indicators, but they are not independent verification of performance.

Tencent says Hy4 will be available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom’s AI Assistant, and other services. APIs are also being made available through platforms including OpenRouter, Hermes, and OpenCode. Pricing remains at about $0.90 per million input tokens and $2.70 per million output tokens. Tencent also plans a lower-cost Hy4 Flash version.

The company says its cybersecurity training emphasizes defensive work, such as identifying vulnerabilities and developing fixes, while excluding offensive exploitation material. Internal reviews also find that Hy4 has strong biological knowledge but more limited capabilities for open-ended experimental design.