AI Agent UpdateTencent launches Hy4: 770B parameters, first Hunyuan model with vision and computer use
Tencent released Hy4 on November 4, its first Hunyuan model with vision and computer use: a 770-billion-parameter system that can read charts and documents and operate software interfaces.
The News
Tencent released the full Hy4 model on November 4, the first in its Hunyuan line with vision and computer use. Hy4 is a 770-billion-parameter mixture-of-experts model, up from 295 billion for Hy3, and it can interpret charts and documents, reason across images, and operate command-line tools and graphical interfaces. In plain terms, it can see a document, understand it, and then do something with software — not just describe it.
The launch is wide rather than limited to one product. Hy4 is live via Tencent Cloud TokenHub, WorkBuddy and CodeBuddy, Yuanbao, ima and WeCom's Dayuan, plus OpenRouter, the marketplace where developers compare and route between models. TokenHub carries a 50% discount through December 31, 2026, and WorkBuddy and CodeBuddy offer a limited free usage quota from November 4 to 15 — a classic launch push to get developers trying the model before they commit.
There are two honest caveats. Tencent says the model weights will be open-sourced soon, but they were not yet out at launch, so independent researchers cannot yet download and test the full model themselves. And a cheaper Hy4 Flash version is promised but undated. The launch is therefore best read as a capability announcement with real availability through Tencent's own products, and full openness still to come.
Tencent's Hy4: 770B parameters, and the first Hunyuan model that can see and act.
Why It Matters
This matters because the model race is shifting from who writes the best paragraph to who can do the most useful work. A model that can read a spreadsheet, understand a chart and then operate software is closer to an employee than a search engine, and that is where businesses will pay. Hy4's combination of vision and computer use puts Tencent in that race with a shipped product, not a research paper.
It matters for the US–China AI balance, too. Western coverage often treats Chinese models as followers, but a 770-billion-parameter multimodal system launched across consumer, workplace and developer channels in one day is a sign of a mature AI organisation. Whether Hy4 matches the best US models on hard benchmarks is still unproven without open weights — but the distribution is not in doubt.
For readers, the practical point is simple: the next generation of AI tools will not just answer you, they will act for you inside software. Hy4 is one more example of that shift arriving as a product, and Tencent's huge existing user base gives it a head start in putting it in front of people. Developers are the audience to watch next: if Hy4 earns real traffic on OpenRouter and in Tencent's own developer tools during the discounted launch window, that usage — not the launch-day headlines — will show whether the model is competitive where it counts, in daily work rather than demonstrations.
Analyst's Take
The verdict is Bullish, for Tencent and for the direction of models generally. Vision plus computer use is the step that turns a model from an answer machine into a worker: reading a chart is useful, but reading a chart and then operating the software to act on it is a product. That Tencent is shipping this at 770 billion parameters, through its own consumer and enterprise products at once, shows how fast Chinese labs are closing the capability gap — not just matching text quality, but shipping the agentic features Western labs are also racing toward.
The caveat is openness. Until the weights are actually released, outside researchers cannot verify Tencent's claims as thoroughly as they could with a fully open model, and the undated Flash version leaves the price-sensitive end of the market waiting. But availability through TokenHub, Yuanbao, WorkBuddy and OpenRouter is real distribution, and a 50% launch discount is a serious attempt to win developers. For Tencent's ADR, the read-through is a stronger AI story attached to products hundreds of millions of people already use.
By The Numbers
FAQ
What is a mixture-of-experts model, in plain terms?
A model made of many specialist sub-models, or experts, where only the relevant experts switch on for each task. That lets the total system be huge — Hy4 has 770 billion parameters, the adjustable values the model learns — without running all of it for every question.
What does computer use mean?
The model can operate software the way a person would: using command-line tools and graphical interfaces, clicking and typing, rather than only answering in text. Combined with vision — understanding images, charts and documents — that turns a chatbot into something that can carry out multi-step tasks.
What is the difference between Hy4 and Hy3?
Scale and senses. Hy3 had 295 billion parameters and no vision or computer use; Hy4 has 770 billion and adds both. It is the first model in Tencent's Hunyuan line that can see and act, not just read and write.
Where can people actually use Hy4?
It is live via Tencent Cloud TokenHub, WorkBuddy and CodeBuddy, Yuanbao, ima and WeCom's Dayuan, plus OpenRouter. TokenHub carries a 50% discount through December 31, 2026, and WorkBuddy and CodeBuddy offer a limited free usage quota from November 4 to 15.
Are the model weights open-source yet?
Not yet. Tencent says the weights — the trained model itself, which others could download and run — will be open-sourced soon, but they were not out at launch. A cheaper Hy4 Flash version is also promised but undated, so the full picture of pricing and openness is still to come.
👁
Source: TechNode / Longbridge