Tencent Unveils AI Model 'Hunyuan4' That Sees the Screen and Operates Software
Tencent has unveiled 'Hunyuan4 (Hy4),' an AI that can look at a screen and operate a computer directly. The Chinese company announced on the 4th (local time) that it had launched the large language model (LLM), which integrates visual understanding and reasoning, computer operation, and long-horizon task execution, marking its full-fledged entry into the global open-source (publicly available technology that anyone can view and modify) AI race. Hy4 demonstrates an evolution beyond simply answering questions or generating code, functioning as an agent (an AI that carries out tasks on a user's behalf) capable of performing work in real software environments.
The key change in Hy4 is the introduction of visual understanding and computer-use capabilities. It can analyze fine details in images, interpret complex charts, and even perform mathematical visual reasoning. It operates both command-line interfaces (CLI) and graphical user interfaces (GUI), automating tasks that humans previously performed on computers. In particular, it received specialized training on the 3D graphics tool Blender and computer-aided design (CAD) software, giving it the ability to plan and execute complex tasks across multiple steps.
Tencent also presented quantified results for real-world enterprise work. In evaluations conducted on Tencent's AI work platform WorkBuddy, Hy4 scored an average of 83.18 points with a 96.15% pass rate, ranking first in nine of 12 evaluation scenarios targeting the Chinese domestic market. When integrated with the collaboration platform WeCom, it comprehensively analyzes chat, email, document, calendar, and meeting data to identify unfinished tasks and approaching deadlines and suggest priorities.
Validation in the financial sector was carried out by practicing professionals. A blind evaluation jointly held with Tencent's fintech division drew a total of 213 financial experts, including 153 experts from external institutions, of whom about 70% had five or more years of experience. Across a total of 579 real-world financial task evaluations, Hy4 scored an overall 3.04 points, ahead of DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68), and GLM-5.3 (2.68). Tencent said the model was trained on thousands of professional work environments and tens of thousands of work cases to deepen its understanding of specialized financial tasks such as fund research, securities analysis, and quantitative investment strategy.
Improvements in reliability were also disclosed. In internal evaluations that strengthened latest-information retrieval, multi-source integration, and fact-checking, the hallucination rate fell from 12.95% in the previous Hy4 preview to 7.2%, while the error rate from unsupported assumptions dropped from 23.1% to 5.6%. In game development, Tencent worked with developers at its Interactive Entertainment Group to improve the model's ability to use major game engines such as Unreal Engine, Unity, and Godot, posting performance gains of 19.1% on a game development benchmark and 9.8% on a game logic evaluation compared with the preview model.
Tencent conducted safety reviews ahead of the model's release. The company plans to publish Hy4's model weights and system card within the next few days. It also plans to expand team collaboration features based on Hy4 through WorkBuddy's 'Projects' feature.
