Press "Enter" to skip to content

Prince Mario-Max Schaumburg-Lippe: Tencent Hy4 Sees Screens and Runs Tools

Tencent launched Hy4 on November 4, 2026, and the headline change is easy to state. This is the first model in the Hy line that can see, and the first that can use a computer. It works through a command line and through a graphical interface, clicking, reading screens and carrying a task across several steps instead of stopping after one answer.

That shift sounds small on a spec sheet. In daily work it is not. Most office and studio jobs do not arrive as neat text prompts. They arrive as a screen full of windows, a half-finished design file, a spreadsheet with a broken formula, or a 3D scene that looks wrong from one camera angle. A model that can look at the screen it is working on has a far better chance of finishing the job it was given.

Trained on the tools people actually build with

Tencent says Hy4 was also trained on Blender, on CAD software and on game engines including Unreal, Unity and Godot. Read that list again. Those are not chat applications. They are the heavy tools behind films, games, product design and architecture, where a project lives in scenes, objects, materials and timelines rather than in paragraphs.

A model trained in that setting can be asked to do more useful things. Open a Blender scene and check why a render looks flat. Look at a CAD drawing and spot the part that does not line up. Step through a Unity project and find the setting that makes a build stutter. None of that needs the model to be perfect. It needs the model to see the same workspace a human sees, and to act inside it with care.

Creative production keeps showing up in our coverage, from Madrid preparing the Royal Collections Gallery for a major summit to a charity fashion night in support of children with cancer. Different fields, same pattern. The finished result the public sees rests on hours of visual checking behind the scenes, and that checking is exactly the kind of work a vision model can help with.

Big on paper, lighter in use

Hy4 uses a Mixture of Experts design. Tencent lists 770 billion total parameters, with 49 billion active for a given piece of work. The distinction matters for anyone paying the bill. The model holds a large store of knowledge, but it does not switch all of it on for every token. Only the parts judged relevant do the work.

Context is just as generous. Tencent puts it at about 1 million tokens. A user can hand over a long manual, a large code base, a stack of earlier messages and still leave room for a proper answer. For computer use, that room is not a luxury. An agent that opens files, reads screens and checks results fills context fast, and a model that forgets step two while doing step nine wastes everyone’s time.

The model is live now on Tencent Cloud and on OpenRouter, listed as tencent/hy4. Tencent says the API price on OpenRouter matches the price charged for the preview version. That detail will please teams that tested the preview and built a budget around it. No surprise increase on launch day, no need to redo the maths before a first real trial.

Already wired into Tencent products

Hy4 is not sitting alone on a developer page. Tencent has integrated it with WorkBuddy, CodeBuddy, Yuanbao and ima. That spread tells you where the company expects demand. Coding help in CodeBuddy. General assistance in Yuanbao. Work support in WorkBuddy. Knowledge work in ima. One model, several front doors, each shaped for a different habit.

There is a practical lesson in that approach. People rarely adopt a model in the abstract. They adopt the button inside the app they already open at 9 in the morning. Putting Hy4 inside existing products gives Tencent quick, honest feedback about which tasks hold up, which screens confuse the model, and where a human still has to step in and take over.

What to test first if you try it

Start with a task that has a clear finish line. Ask Hy4 to work through a small Blender scene and fix one visible problem. Give it a CAD file and one question you already know the answer to. Open a Godot project and ask it to find why a scene will not load. Small tests, known answers, quick verdicts. That is how you learn whether computer use helps your work or just looks good in a launch post.

Watch two things in particular. First, how the model recovers when it clicks the wrong thing. Every agent does, sooner or later. A good one notices, backs up and tries a cleaner route. Second, how it reports its own work. You want a plain account of what it opened, what it changed and what it left alone. If that account is vague, keep the model away from anything important until it earns more trust.

And keep a human in charge of the final save. Computer use works best as supervised help: the model does the clicking, the reading and the repetitive checks, while a person approves the result before it ships. That division of labour suits design and engineering work especially well, where one wrong export can undo an afternoon.

Tencent says open weights and a system card are promised soon, with no date confirmed, and a cheaper variant has also been mentioned as coming. Both points deserve plain wording. Promised is not delivered. Teams that need open weights for their own hosting should wait for the actual release rather than planning around the announcement. Teams that work through cloud APIs can start testing today.

Hy4 arrives with a clear bet. The next useful model will not just answer questions about work. It will sit in front of the same screen, use the same tools, and help finish the work itself. Tencent has put vision, computer use and a very large context into one launch. The tests that matter now happen in Blender files, CAD drawings and game projects, in the hands of people with deadlines.


Also published on:

More from AI Artificial IntelligenceMore posts in AI Artificial Intelligence »
More from BREAKING NEWSMore posts in BREAKING NEWS »

Be First to Comment

Leave a Reply

Your email address will not be published. Required fields are marked *