JOURNAL / 2026.11.08
Tencent Hy4 adds vision and computer use: what changes for agents
Hy4 broadens the work Tencent's agents can attempt. Documentation makes access and image limits more concrete, but evaluating a complete task requires more than recognizing a screen.
A desktop application can show that something went wrong without providing an error another program can read. For an agent, recognizing that state and deciding how to fix it opens a different path from working with text alone. This possibility makes Hy4, introduced by Tencent on November 4, relevant: the company announces visual understanding and computer use, with specialized training on tools such as Blender and CAD.
The starting point was already ambitious. Hy4 Preview, released in August, targeted coding, office work and research, with a context exceeding one million tokens. The interest of the current release therefore lies in broadening how it receives information and acts. A large context might contain an entire project while still missing a decisive signal that appears only in the interface.
The official API guide makes part of the change concrete. It distinguishes the two versions' identifiers and assigns them the same nominal limits: one million tokens of context, 960,000 for input and 64,000 for output. It documents images mixed with text and conversations containing multiple images. It also warns that images are ultimately resized to approximately 2048 × 2048 pixels: sending a larger file does not guarantee the model sees more detail.
That limit has practical consequences. Consider a spreadsheet with narrow columns: getting the overall appearance of the window right is not enough to distinguish a number or a unit. I would test whether it could read those cells before allowing the agent to change a formula. A crop might preserve relevant detail, although it could also remove the context explaining which table the agent is viewing. The integration will have to address both problems; accepting an image alone does not solve them.
From a screenshot to a completed task
OpenRouter lists Hy4 with text and image input, text output and tool calling. This is a documented hosted offering, not a test we have run. The distinction also clarifies the model's role: operating an application requires a system that turns its decisions into actions, captures what happened and returns evidence of the new state.
In a hypothetical design workflow, I would ask it to modify a part and save the result in a particular format. There are several opportunities for failure: selecting the wrong object, entering a measurement in a different unit or saving a screenshot instead of the editable file. A convincing visual demonstration could hide any of them. My success criterion would be reopening the file and checking its properties. This evaluates what the user needed, without confusing a plausible sequence of clicks with a completed assignment.
The Preview model card acknowledged reasoning longer than necessary and a tendency to over-verify. I would not automatically attribute those faults to Hy4. I would retain the question they raise: how much additional work the pursuit of certainty creates. In a graphical application, each inspection can add another screenshot and another interaction; better verification need not mean repeating the same observation more often.
A relevant distinction remains between access and openness. The November announcement promises weights and system cards in the following days. We have not confirmed the final version's artifacts here. Preview's weights are not enough to attribute the same license or self-hosting possibilities to Hy4.
My reading favors extending agents into applications without convenient integrations. It could make work that is currently difficult to automate accessible without rebuilding the tool. But the evaluation that interests me most now is a comparison between Preview and Hy4 on the same assignments, in the same environment, with an outcome verifiable outside the conversation. I would measure accepted tasks, human intervention and total cost. If vision reduces the corrections a person must make, we will have a useful improvement; if it merely produces a more convincing explanation of a screen, the gain will be considerably smaller.
Primary sources consulted on November 8, 2026. We have not run Hy4 or reproduced its evaluations. The examples and verification criteria are our own proposals.