What Is Tencent Hy4? Vision, Computer Use, and the Open-Weights Question

Tencent Hy4 is Tencent’s latest flagship artificial intelligence model, released in full on 4 November 2026. It is the successor to the Hy4 Preview from August 2026, and it is the first model in Tencent’s Hunyuan family that can both understand images and operate a computer. In plain terms: earlier Hunyuan models could only read and write text, while Hy4 can look at a chart, a screenshot or a diagram — and then take actions inside software to get a job done.

This guide explains what Hy4 is, how its vision and computer-use abilities work, where you can try it, what it costs, and the one important qualification on its “open-source” label.

Hy4 in one paragraph

Hy4 is a large language model built by Tencent, the Chinese technology company behind WeChat, QQ and a large gaming business. It is a mixture-of-experts system with 770 billion parameters in total, of which about 49 billion are active for any single task. It accepts around one million tokens of context in a single conversation — enough for several long books’ worth of text, or a large codebase, at once. It is available through Tencent’s own products and cloud platform, and through third-party services such as OpenRouter. Tencent says its pricing is unchanged from the preview version, and two of its products are free to use with Hy4 until 15 November 2026.

What “mixture of experts” means

A mixture-of-experts model is not one giant network that runs in full for every question. Instead, it contains many smaller specialist networks — the “experts” — plus a routing mechanism that picks a subset of them for each piece of input. Hy4 has 770 billion parameters in total, but only around 49 billion are switched on per token. The point is efficiency: the model can hold far more total knowledge than a dense model of the same running cost, because most of its capacity stays idle on any given request. The trade-off is size on disk — all those experts still have to be stored somewhere — which is why downloading a model of this scale takes over a terabyte of storage, and why running it locally needs a cluster of high-end graphics cards rather than a single machine.

For ordinary users this detail matters only in one way: it is how Tencent can offer a very large model at a low price. Hy4’s listed API pricing is about $0.90 per million input tokens and $2.70 per million output tokens, with cached input at about $0.045 per million — cheaper than most closed frontier models, and the same rates as the preview despite the new capabilities.

Vision: what Hy4 can see

Vision is the first of Hy4’s two genuinely new abilities. The preview version was text-only; the full release adds native visual understanding, meaning the model processes images directly rather than relying on a separate captioning step. Tencent describes three kinds of visual work it can now handle:

• Fine-grained image reading — picking out small details in a photograph or screenshot, not just the overall subject.
• Complex chart analysis — reading values, trends and labels from charts and graphs, the kind of visual information that fills business reports and scientific papers.
• Mathematical visual reasoning — working through problems presented as images, such as a handwritten equation or a geometry diagram.

Why does this matter? Because a large share of knowledge work is visual. A doctor’s scan, an engineer’s drawing, a sales chart, a photo of a broken part — all of these carry information that is awkward to describe in words. A text-only model needs a human to translate the picture into text first, and that translation loses detail. A model with vision skips the middleman. It still makes mistakes — reading a chart is not the same as understanding it — but the capability changes which tasks can even be attempted.

Computer use: from answering questions to doing jobs

The second new ability is the more consequential one. Computer use means Hy4 can interact with software the way a person does: typing into a command line, clicking through a graphical interface, moving between applications, and checking its own work. Tencent says Hy4 received specialised training on professional software including Blender, used for 3D creation, and CAD tools used in engineering and design, and that it is designed for long-horizon tasks — jobs that take many steps and where an early mistake compounds if it is not caught.

Think of the difference this way. A chatbot answers the question “how do I build this spreadsheet?” Computer use builds the spreadsheet, notices a formula error, fixes it, and reports back. That is the direction the entire frontier is moving: Anthropic’s Claude, OpenAI’s agent features in ChatGPT, and Google’s Gemini have all added versions of this ability, and Hy4 is Tencent’s entry into the same race. The practical test is not whether the model can click buttons — it is whether it can recover when something unexpected happens, like a dialog box it has never seen, without a human stepping in. Tencent’s claims on this front rest on its own evaluations so far, and independent testing will be the real verdict.

Inside Tencent’s products, computer use shows up as workplace assistance. In WorkBuddy, Tencent’s office collaboration product, Hy4 connects to chats, emails, documents, calendars and meetings, tracks project history, flags upcoming meetings that need preparation, and watches approaching deadlines. In financial analysis, Tencent trained the model on thousands of specialised workspaces and tens of thousands of realistic tasks covering fund research, securities analysis and quantitative investment work. In software and game development, it is tuned for complex business logic and for engines including Unreal, Unity and Godot.

How good is it? Reading the numbers honestly

Tencent’s announcement includes several benchmark-style claims, and they need careful reading. In a blind evaluation run with Tencent Financial Technology, 213 financial professionals judged 579 real-world tasks, and Hy4 scored 3.04 overall, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68. In an internal investment-research benchmark, Hy4 improved 12.03 percentage points over the preview. And in internal evaluations, its hallucination rate fell from 12.95% in the preview to 7.2%, while errors from unsupported assumptions fell from 23.1% to 5.6%.

All of these are Tencent’s own measurements. That does not make them false — internal evaluations are how labs steer development — but it does mean they are not yet independent evidence. Until outside researchers and benchmark organisations run the full model, the honest summary is: Tencent reports a large jump over its own preview, and claims the top spot among open-weight models on its chosen financial evaluation. Treat leaderboard claims as provisional, and watch for independent results in the weeks after release.

Where to try Hy4, and what it costs

Hy4 is available now in several places:

• Tencent products: WorkBuddy, CodeBuddy, Yuanbao, the WeCom assistant Dayuan, and ima.
• Cloud API: Tencent Cloud TokenHub.
• Third-party platforms: OpenRouter, Hermes and OpenCode, with more integrations underway for products including QQ and Tencent News.

Pricing is unchanged from the preview at roughly $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens. Two offers soften the cost further: WorkBuddy and CodeBuddy are free from 4 to 15 November 2026, and Hy4 on TokenHub is half price through 31 December 2026. A cheaper, faster variant called Hy4 Flash is promised soon for users who prioritise cost over maximum capability, but it has no release date yet.

The open-weights question

Here is the qualification that matters most. Tencent describes Hy4 as an open-source model, and the Hy4 Preview was indeed open-sourced in August 2026. But for this full version, Tencent says only that it plans to release the model weights and a comprehensive system card “in the coming days”. As of the announcement, those files are not public.

Why does this distinction matter? “Open weights” means anyone can download the trained model, run it on their own hardware, inspect it, and build on it without depending on the vendor’s servers. A hosted API, however cheap, is not the same thing: the vendor can change it, restrict it, or withdraw it, and researchers cannot verify claims against the actual model. Until the weights appear, Hy4 is best described as a hosted release with an open release promised — not yet a downloadable model. If the weights arrive within days as promised, Hy4 will become one of the most capable openly available models in the world. If they slip, the announcement will deserve a more sceptical reading.

Safety: what Tencent says it built in

Tencent’s announcement gives unusual detail on safety, and it is worth summarising because it shapes what the model will and will not do. On cybersecurity, Hy4 was trained on defensive data — proof-of-concept development and vulnerability remediation — while offensive material about exploitation was explicitly excluded from training. Tencent says its Xuanwu Lab and Keen Security Lab assessed the model across cryptographic reverse engineering, binary deobfuscation and vulnerability exploitation, and found no significant concerns about offensive capability in the areas tested. On biology, the company says Hy4 has strong foundational knowledge but is more limited than frontier rivals at open-ended experimental design. These are the company’s own assessments, not outside audits, but the specifics — what was excluded, who tested it — are more transparent than a generic safety statement.

Hy4 in the wider open-model landscape

Hy4 arrives in the middle of a fast-moving open-weight wave. In the weeks around its release, Moonshot AI published the weights for Kimi K3.1 Pro, a 2.8-trillion-parameter model, and Mistral opened the weights for Mistral Large 4. The pattern is consistent: Chinese and European labs are competing not just on benchmark scores but on openness, while the leading American labs keep their most capable models closed. Hy4’s pitch is that it combines frontier-class agent abilities — vision plus computer use — with open-model pricing and, soon, open weights. Whether that combination holds depends on the weights actually arriving and on independent benchmarks confirming the agent claims.

What to watch next

• The weights and system card: promised within days of the 4 November announcement. Their arrival turns Hy4 from a hosted product into a genuinely open model.
• Hy4 Flash: the cheaper, faster variant, for users who want the capabilities at lower cost.
• Independent benchmarks: outside labs testing vision, computer use and long-horizon tasks against Claude, GPT and Gemini-class rivals.
• Everyday agent reliability: the real test of computer use is not demos but weeks of ordinary work — recovering from errors, asking for clarification, and knowing when to stop and ask a human.

Hy4 is not just a bigger chatbot. It is Tencent’s argument that the next stage of AI is models that perceive the screen and act on it, available at open-model prices. The vision and computer-use capabilities are real and available today through Tencent’s products and APIs. The open part is, for now, a promise with a short deadline — and that deadline is the most interesting thing about this release.

For the day-by-day news version of this story, see our report: Tencent’s Hy4 Learns to See and Operate a Computer — Weights Still to Come. Tencent’s own announcement is the primary source for this guide: Tencent launches Hy4.