Tencent's new Hy4 can understand images and operate software on its own

Chinese tech group Tencent launched the full version of the Hy4 AI model on November 4, 2026. It replaces a preview version from August, which Tencent called Hy4 preview.

The big news is that the model can now understand images. According to Tencent, it can interpret charts and documents and reason about relationships across multiple images. A few days earlier, on October 29, Tencent senior executive vice president Dowson Tong had signaled that image understanding was on the way.

Can operate software

Hy4 can also control computer programs. That includes both text-based command-line tools and programs with ordinary buttons and menus. Tencent has trained the model specifically on programs such as Blender and CAD, which are used for 3D graphics and design, and on tasks that span multiple programs.

The training targets game development in tools such as Unreal Engine, Unity and Godot, as well as finance work. The model is also meant to be able to plan long tasks across multiple applications. According to TechNode, development has prioritized office tasks, financial analysis and game development.

Tencent's own figures

Tencent says hallucinations, meaning cases where the model makes things up, have fallen from 15.76 to 5.45 percent. Speculative errors have fallen from 23.1 to 5.6 percent. Tencent also reports a 19.1 percent improvement on GameDev and 9.8 percent on GameLogic compared with the preview.

In a finance evaluation with 213 professionals and 579 tasks, Hy4 ranked first among open weight models with a score of 3.04. The figures come from Tencent and have not been independently verified. It is unclear what some of the figures, such as the hallucination rates, are based on.

Tencent itself acknowledges that there is still a gap to the best closed models, especially in advanced math and advanced visual question answering. On the Hy-BioExperimentBench test, Hy4 scored 30.4 out of 100, which shows a limited ability to turn biological knowledge into protocols. Cybersecurity data is limited to defensive purposes, and material on exploit development and penetration testing has been filtered out.

Pricing and access

API pricing is $0.834 per million input tokens and $2.501 per million output tokens, unchanged from the preview. Tokens are the small chunks of text AI models work in. Tencent Cloud TokenHub offers a 50 percent discount through December 31, and there is a limited free quota from November 4 to 15.

The model is available through Tencent Cloud and OpenRouter, as well as in Tencent's own products CodeBuddy, WorkBuddy, Yuanbao and ima. The context window, meaning how much text it can handle at a time, is about one million tokens (1,048,576) on OpenRouter.

The preview version is open source, but the model weights for the full Hy4 have not yet been released. They and a system card are planned for later. Tencent is also planning a cheaper and faster version, Hy4 flash, which is meant to retain image understanding and the ability to carry out tasks on its own.