On the morning of November 4, at Tencent's booth at the Light of the Internet Expo in Wuzhen, the screen still showed the preview build, but the talk was about the official version released just hours earlier that morning.
That day, Tencent Hunyuan released Hy4, its next-generation large language model. The line in the official announcement that stood out most was that it could, for the first time, operate a computer on its own: move the pointer, find a button, click it, with no one taking over in between.
It sounds like advertising. Take it apart, though, and it is a plain enough thing. For two years, large models have been good at talking. You ask, they answer. You hand them a paragraph, they rewrite it. They have always been behind a pane of glass.
That pane has now been cracked open a little. Hy4 supports what the industry calls computer use, working through both a command line and a graphical interface. The command line is the black screen with white text; the graphical interface is the windows, buttons and menus we see every day.
Mixing the two means one thing: it can see what is on the screen, and it can decide where to click. To the model, a computer is no longer a pile of text. It is an object it can act on.
The Parameter Count Climbs Again
Parameter scale is the other number mentioned again and again in this release. Tencent says the previous Hy3 series came in under 300 billion parameters, while the Hy4 series reached 770 billion. The next generation, Hy5, will grow further.
That figure is not the largest in the industry. But Tencent's rhythm is one major version every two months: open a preview first, collect feedback, then ship the official release. Since rebuilding its infrastructure in February this year, the company has kept to that road for nearly a year.
What does 770 billion parameters feel like to an ordinary person? It is less like a library and more like its index. Not the books themselves, but the thing that decides how fast you can find one. A finer index also costs more power to page through.
A bigger model directly lifts basics such as long-document comprehension and instruction following. Tencent says Hy4 is noticeably better than the preview on these fronts. There is no third-party evaluation behind that claim, so it is worth filing away rather than repeating.
Eyes for the First Time
More interesting than the parameter count is the seeing. Hy4 introduces visual understanding for the first time: it can handle fine-grained images, analyze complex charts, and do math problems that arrive as pictures. For a model that could only read text, this is a new set of senses.
What does that mean in daily life? A simple example: an expense receipt photographed badly, with someone else's handwriting pressed into the corner. Older models could only read what you typed. This one can first make out the image.
Chart analysis is another direction. An industry report runs dozens of pages, full of line charts, bars and pies. A person scans it and grabs the point; a machine used to read only the captions underneath. Now it can see where the line bends.
Seeing and acting together are what close the loop: see the screen, understand the screen, then do something on the screen. Take away either half, and the so-called agent is just a chat window.
Professional Software Is the Hard Part
Tencent specifically mentions extra training on professional software such as Blender and CAD. Those names mean little to most people, but they are everyday tools in many trades: animation, industrial design, architectural drafting.
These programs share a trait: many buttons, deep menus, an operation that takes several steps, and dialog boxes that stop to ask for parameters. Unlike a web page, they offer no uniform interface that a machine can simply call.
That is why "can operate a computer" is hardest to deliver here. Opening a browser and searching for something is not remotely the same difficulty as changing a chamfer inside modeling software. Every step of the latter requires watching the screen and judging.
Tencent's account is that after the extra training, the model's planning and execution improved on these long, multi-application tasks. That conclusion currently rests on the company's own word, with no third-party reproduction, so it is worth noting and no more.
Open Source, Not Yet Open
On launch day, the easiest phrase to misread was open source. The precise version: Hy4 will open-source its model weights in the near future, along with a model system card. The weights are not out yet, and there is no file to download.
The difference matters. Only with weights released can others run, modify and distill the model on their own machines. Until then, it is a service reachable through an interface, with the price set by the provider.
The system card is a separate document. It covers the model's capability boundaries, the rough makeup of its training data, and known risks. Tencent says it will focus on network and biosecurity risks, which is a careful way of putting it.
The company also says this is the first time Hunyuan has ranked in the top tier of open-source models. Note the phrase "the company says." In the same breath, the Hunyuan team admits there is still a considerable gap with the top closed-source models.
Read the two sentences together and they are truer than either alone: it has reached the front of the open-source tier, but that tier itself is still some distance from the very front, and no single release will close that distance.
The Arithmetic on the Price Sheet
Pricing carries over from the preview: 6 yuan per million input tokens, 18 yuan per million output tokens, 0.3 yuan for a cache hit. During the launch period, Tencent Cloud charges half the list price, a discount that runs to the end of December. What happens after that has not been said.
That is not expensive by domestic standards. But the real cost is not the unit price; it is how many steps the model has to run. A task that requires twenty mouse clicks multiplies input and output many times over, and the bill has to be calculated across the whole workflow.
It is already wired into Tencent Cloud, Yuanbao, the WeCom assistant and several office products. Tencent properties such as QQ, Honor of Kings and WeRead are being connected as well, spreading faster than the model itself iterates.
For ordinary users, the first thing they may notice is probably not that it can operate a computer. It may be an afternoon when you ask it, inside WeCom, to stitch several group chats, a few emails and a calendar into a weekly report.
How well it does that will be known only when more people actually use it. The demo on stage was smooth. Real windows lag, pop up dialogs, demand captchas, and lose focus the moment you click away.
A hand that can click a mouse is still far from doing your job for you. But it has at least let the model out of the chat box and onto the screen we stare at for hours every day. The step is small; the direction has changed.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.