Tencent Launches Language Model Hy4 with Computer Control Capabilities

On November 4, 2026, Tencent unveiled the full version of the Hy4 language model. The model has learned to analyze images and operate a computer through the command line and graphical interface. It is designed for multi-step tasks in applications.

What's New in Hy4

The model can analyze images, complex graphs, and visual mathematical tasks. For interacting with a computer, Hy4 combines screen content recognition, reasoning, and tool invocation. This allows it to perform sequences of actions in applications.

According to Tencent, the model underwent separate training for working in Blender and computer-aided design (CAD) programs. Developers also trained the model for long-term tasks in multiple applications. However, the presence of these features does not mean that Hy4 flawlessly handles any project.

What Results Tencent Claims

The company reported improvements in programming, document handling, scientific research, and financial analysis. In financial analysis, Tencent and a partner involved 213 specialists and 579 tasks for evaluation. Hy4 scored 3.04 points and outperformed several models in the selected comparison. The result cannot be considered an independent assessment of all the system's capabilities.

Tencent opened the preliminary version Hy4 Preview on August 28. The new version expands its functionality, primarily through visual perception and program management. The company has not presented direct evidence of accelerated computations or guaranteed user productivity growth.

Where the Model is Available and When Weights Will Be Released

Hy4 is already available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, and other company services. The API is also accessible via OpenRouter. Cloud access does not allow independent deployment of the model.

Tencent promises to publish Hy4 weights and system cards with information on tests and safety measures in the coming days. As of November 5, the open weights of the full version have not yet been presented. The company separately reported limiting training materials related to offensive cyberattacks, but independent verification of the claimed measures is still limited.

The practical outcome of the launch will depend on how reliably Hy4 performs tasks in real interfaces and what usage conditions the developer sets after the publication of the weights.