Tencent Releases Hy4 AI Model with Computer Control and GUI Support

On November 4, Tencent officially released the final version of Hy4, the flagship model within the Hunyan family. For the first time in this product line, the model has been trained to analyze and understand images, as well as interact with computers through both graphical user interfaces and command line interfaces.

• Hy4 introduces the ability to operate computers via GUI and CLI, enabling interaction with complex software.
• The preliminary version was launched on August 28, followed by the full release just over two months later.
• The model requires 770 billion parameters in total, with 49 billion active parameters.
• Hy4 is already available for integration and testing through the TokenHub cloud API.

Advanced Capabilities and Computer Control

The standout feature of the Hy4 model is its capability to autonomously interact with operating systems and applications. The system analyzes screen contents, determines necessary actions to achieve goals, executes them, and evaluates the final results.

This functionality allows Hy4 to work effectively with demanding professional applications such as Blender and various CAD systems. The developers focused heavily on multimodality and multi-step task execution to position the model competitively in the AI agent landscape.

Performance Benchmarks and Safety

Compared to the preliminary release, performance metrics have improved notably. GameDev benchmark scores increased by 19.1%, GameLogic by 9.8%, and the iRaB financial test score rose by 12.03 percentage points, while factual error rates dropped from 12.95% to 7.2%.

Regarding cybersecurity, training was intentionally limited to defensive scenarios focused on vulnerability detection and patching. This aligns with a growing trend among leading AI developers to implement stricter safety measures on advanced models.

Implications for Developers and the Market

The introduction of Hy4 highlights Tencent's drive to keep pace with global AI advancements in autonomous desktop agents. While coding benchmarks show mixed results, strong scores in environments like OSWorld 2.1 demonstrate significant potential.

Deploying the model locally requires robust server infrastructure due to its size of 770 billion parameters with 49 billion active weights. However, cloud-based API access provides a practical alternative for developers looking to evaluate its capabilities.