Hy4 is not a chatbot upgrade. It is the first Tencent model that can look at a screen and then work it.
The final Tencent Hy4 build landed on November 4, as the finished version of the Hunyuan Hy4 preview from August. That preview shipped without vision. This one reads charts, parses documents, solves maths problems from an image, and drives a desktop through both the graphical interface and the command line.
That last part is the reason to pay attention. A model that clicks through application windows is a different product. And it answers with actions, not descriptions.
What is actually new
Most of the release is scale. A few pieces are capability.
- Size: 770B total parameters, mixture of experts, 49B active per token.
- Context: up to roughly one million tokens on Tencent’s own endpoint, with about 960K input and 64K output.
- Vision: first time in this line. Chart reading, multi-image comparison, document parsing, visual maths.
- Computer use: graphical interface control plus command line, tuned for CAD and Blender workflows.
- Engines: adapted for Unreal, Unity, and Godot, with GameDev scores up 19.1% over the preview.
And the vision piece is table stakes in 2026. The computer use piece is not.
Hy4 computer use is the piece that changes the product. Tencent claims cross-application task execution. That means the model holds a goal across several programs. It does not complete one step and stop.
The numbers a buyer cares about
Tencent kept the preview pricing. But that is unusual, since parameter counts went up sharply.
Cost per task is the number that decides whether an agent product is viable. A single agent run can make dozens of model calls. At $0.78 per completed task, a workflow that fails half the time still costs less than a dollar to attempt. At the rates charged by the top closed models, the same retry loop is a budget line.
Input is ¥6 per million tokens. Output is ¥18 per million. Cached input drops to ¥0.3 per million, which is the line that decides whether an application is affordable at volume.
Two discounts sit on top. Tencent Cloud’s TokenHub endpoint is at half price through December 31, 2026. WorkBuddy and CodeBuddy carry a free allowance from November 4 to November 15.
Artificial Analysis puts the cost per task at about $0.78. That figure is the one to hold onto, because it is measured against the same benchmark where Hy4 ranks second among Chinese models, behind Kimi K3.1 Pro and level with Grok 4.7 on its high setting.
Two months between versions
Tencent rebuilt its compute base in February. It has shipped a new version roughly every two months since. The pattern is preview first, feedback, then general release.
Hy4 preview arrived in late August. The finished build landed on November 4, a little over two months later, with vision and computer use added in between. That is a fast loop for a 770B model, and Western labs have not matched it on open releases.
It also means the version you evaluate today has a successor. That successor already has a rough date. Planning a product around this tier means planning for a model swap every quarter.
Where the rankings actually land
Two leaderboards matter here, and they say different things.
On the Artificial Analysis intelligence index, Hy4 sits second among Chinese models across 724 entries. On Agent Arena it ranks 13th in the table and first among open models. So the second number is the more useful one for anyone building agents. That board tests multi-step task completion rather than single-turn answers.
Tencent’s own numbers are narrower. An internal office benchmark scored 83.18 with a 96.15% task pass rate, and a financial test run with 213 outside experts also put it first among open models.
What “open” still means here
Tencent says the weights will be released soon, alongside a system card. As of this week, they are not out. The model is available through Tencent Cloud, WorkBuddy, CodeBuddy, Yuanbao, WeCom’s assistant, and third-party routers including OpenRouter.
So the model is accessible without being open. Anyone who needs to run it on their own hardware is waiting.
The release also carries explicit safety conditions. Training data was restricted to defensive security material, and Tencent says two internal audits found no strong offensive cyber capability.
Those constraints will shape the open weights. A model card with data restrictions attached is a different artefact from a raw checkpoint.
The interesting claim is not the 770B. It is that a model at this price can now operate a desktop. That shifts what a small team can automate, and it is the part competitors will have to answer.
