Tencent's new Hy4 can see a screen and use a computer, but its open part is still to come

Cover image: AI-generated

Hi everyone,

Yesterday, Tencent released the full version of its Hy4 model, and the new part is easy to say in plain words. It can now see images, and it can use a computer.

The August preview could only work with text. This full release adds vision and computer use, so Hy4 can read fine details in an image, make sense of a complex chart, and work across command line and graphical interfaces. Tencent also trained it for hands-on work in tools like Blender and CAD, and for game work in Unreal Engine, Unity, and Godot.

For a regular user, here is what that means in practice.

First, it can look at something, not just read words about it. A chart, a screenshot, or a design can go in as an input.

Second, it can take steps inside software, not just answer a question and stop. That is the part that makes it an agent model rather than a chatbot.

Third, and this is the part to watch, the open part is not here yet. The preview arrived in August with its weights open, but the full release is live through hosted APIs and Tencent products like WorkBuddy, Yuanbao, CodeBuddy, and ima, while the weights and a full system card are still promised for the coming days. The licence for the full release has not been stated yet either.

Tencent also says a lighter version, Hy4 Flash, is coming next, aimed at lower cost and faster work. The listed API price for the full model is unchanged from the preview.

My own view is simple. A release should be judged by what you can actually check and use, not only by what the announcement says. Vision and computer use are real steps forward, and they matter most when the model has to work inside real apps. But calling a model open while its weights are still to come asks for a little patience, and patience is fair here.

If you have tried a model that works inside your apps, what did it get right, and where did you stop and take the next click yourself?