Tencent wants its latest AI model to do more than answer questions. It wants the system to operate the software people use every day.
The Chinese technology company introduced the full version of Hy4 on Wednesday, expanding its earlier text-only preview with native vision and computer-use capabilities. The 770-billion-parameter mixture-of-experts model is designed to interpret visual information, navigate command-line interfaces, and interact with desktop applications, including Blender and computer-aided design software.
The release reflects a broader push toward AI agents that can carry out tasks across software environments rather than simply generate responses. For developers and businesses, the question is whether those capabilities are reliable enough for real-world workflows.
Hy4’s architecture features a 1-million-token context window, allowing it to process massive code repositories and complex operational histories. Rather than chasing generic benchmarks, Tencent trained the model alongside internal business units, targeting complex production environments.
Through collaboration with its Interactive Entertainment Group, Tencent adapted Hy4 for game engines such as Unreal Engine, Unity, and Godot, reporting a 19.1% gain on internal game development benchmarks over the preview version.
In financial services, Tencent evaluated Hy4 across 579 tasks with 213 industry professionals. The company said the model scored 3.04, leading rival open-weight offerings like DeepSeek-V4.1-Flash (2.97) and Kimi K3.1 Pro (2.68).
Tencent also acknowledged its limits. The company noted that Hy4 still trails top proprietary systems in advanced visual question answering, high-level mathematical construction, and long-horizon software engineering.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
The agentic disconnect
While consumer artificial intelligence has settled into conversational search and text drafting, Hy4 reflects a deliberate shift toward agentic software manipulation. Western developers have leaned heavily toward closed ecosystems like Anthropic’s Claude for automated user-interface control.
Tencent’s push to bring computer use into an open-weight framework creates an alternative route for enterprises that want workflow automation without piping sensitive screen data to external proprietary cloud hosts.
However, operating software visually remains unpredictable. Unlike direct code execution or structured API integration, graphical user interface automation depends heavily on screen layout stability. Small visual shifts, rendering delays, or unexpected popups frequently derail visual agents.
While targeted training in Blender and CAD demonstrates ambition, production viability will hinge on whether the model can handle interface errors autonomously without requiring constant human intervention.
Deployment footprint for everyday work
Tencent is positioning Hy4 for workplace automation, software development, and creative applications, with several tools and services intended to make its capabilities accessible to businesses and individual users.
- Workplace automation: Integrated into WorkBuddy, CodeBuddy, and WeChat Work (WeCom), Hy4 synthesizes chats, schedules, and emails to manage action items and track deliverables across active projects.
- Creative and visual tasks: End users gain tools to generate functional frontend interfaces and 3D scenes directly from text prompts, reducing manual UI layout setup.
- Pricing and enterprise control: Listed at $0.834 per million input tokens and $2.501 per million output tokens, Hy4 is available through Tencent Cloud TokenHub and third-party aggregators including OpenRouter.
Tencent is offering a limited free-access period through WorkBuddy until Nov. 15, while the model’s full weights are expected to become publicly available later.
For enterprises considering computer-use agents, the more important test will be how reliably Hy4 performs outside controlled demonstrations. Its ability to handle unexpected interface changes, recover from failed actions, and operate within existing security controls will determine whether the technology becomes useful for everyday work.
In other news, Google is rolling out a fix for Android Auto connectivity issues affecting Samsung Galaxy phones running One UI 9, following reports of failed connections, stuttering, and unexpected disconnections, although the company has not confirmed when the rollout will be completed.