Tencent's Hy4 can now see images and operate a computer

Hy4 is Tencent's flagship AI model, the engine behind its chat and work apps. It is a large language model, which means software trained on a mountain of text to predict and write useful answers.

The numbers tell you how big it is. Hy4 has 770 billion parameters, the adjustable numbers a model learns during training. It also has a context window, which means the amount of text it can keep in mind at once. Hy4's is 1 million tokens. A token is a small chunk of text, roughly three quarters of a word. One million tokens is a stack of long books.

Hy4 uses a design called mixture-of-experts. Picture a hospital where every patient sees only the right specialist, not every doctor in the building.

Only 49 billion of the 770 billion parameters work on any one request. That keeps the cost of running the model, called inference, much lower than a model that uses everything every time.

Tencent released a preview on August 28 and open-sourced it. The full version adds two big abilities.

• Vision: it can read fine details in images and understand complex charts, and it can reason across several images at once.
• Computer use: it can operate command-line tools, where you type instructions, and graphical programs, where you click buttons.
• Targeted training: Tencent trained it on Blender, a 3D design program, and on CAD software, which engineers use to draw parts.
• Multi-app jobs: it is meant for tasks that jump between several programs.

Tencent also says it improved instruction following, long-document understanding and agent tasks. An agent task is a job where the AI plans steps and uses tools on its own.

In real life Say you have a photo of a sales chart and a spreadsheet. A model with these skills can read the chart, open the spreadsheet app and build a summary page, without you describing each click.

Tencent says yes. A hallucination is when an AI states something that is not true. Tencent reports that Hy4's hallucination rate fell from 12.95% in the preview to 7.2%.

Two more internal numbers back that up:

• Unsupported assumptions: errors caused by guessing fell from 23.1% to 5.6%.
• Search answers: hallucinations in Tencent's internal search tests fell from 15.76% to 5.45%.

The catch: these are Tencent's own tests. A benchmark is a standard exam used to compare models, and when the maker writes the exam, outside testing is still worth waiting for.

Tencent pushed hard on practical jobs. Game development scores rose by 19.1% on one benchmark and 9.8% on another, compared with the preview. The model supports work in Unreal Engine, Unity and Godot, three popular tools for making games.

For finance, Tencent ran a blind test with 213 financial professionals and 579 real tasks. Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97, and Kimi K3.1 Pro and GLM-5.3, which both scored 2.68. Blind means the judges did not know which model wrote which answer.

On WorkBuddy, Tencent's own office assistant app, Hy4 averaged 83.18 and passed 96.15% of tasks.

Tencent says the preview learned from high-quality data built together with its own experts. The fields were software engineering, gaming, finance and security, among others. The model was also tuned alongside Tencent's apps, WorkBuddy and CodeBuddy, so the training matched what people actually ask those apps to do.

That explains the product focus. Hy4 is not pitched as a chatty assistant. Tencent describes it as a model built for productivity, meaning real work like fixing code, building reports and making game prototypes.

Most AI models live inside a chat box. You type, they type back. Computer use breaks that wall: the model looks at what is on the screen and decides what to click or type next.

That matters because a lot of office work lives in programs with no easy plug-in. A model that can read a screen can work with those programs as they are. It also raises the stakes for mistakes, which is why the lower hallucination rate is the number to watch.

In game development, the preview could turn a single plain-language request into a playable prototype. Developers could then keep refining a bigger project over many turns of conversation.

In science, Tencent said the preview improved on research problems in AI development, molecular dynamics simulation (computer models of how molecules move), condensed-matter physics and basic mathematics. Tencent also said the preview studied the slow spots in its own serving system and made several rounds of fixes, such as merging small steps into one.

The preview also came with an internal blind test. 163 experts judged 203 engineering tasks, and the preview scored 2.99 out of 4.00. That was slightly ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94.

Tencent also said the preview helped build itself. It proposed training changes, ran experiments and learned from the results, which Tencent called an early-stage self-improvement loop. The company expects the full Hy4 to carry that work forward.

On October 29, Tencent executive Dowson Tong said at the company's Global Digital Ecosystem Conference that Hy4 and Hy4 Flash would support visual understanding. The launch delivers the first half of that promise.

Prices did not move from the preview. Depending on the exchange rate used, Tencent's API charges about $0.83 to $0.90 per million input tokens and $2.50 to $2.70 per million output tokens. An API is a doorway that lets other apps plug into a model.

• Cached input: about $0.04 per million tokens, for text the model has already seen.
• Discount: 50% off on Tencent Cloud TokenHub through December 31.
• Free trial: limited free use in WorkBuddy and CodeBuddy from November 4 to 15.
• OpenRouter: the model has a 1,048,576 token window there and a 65,536 token limit on each reply.

Hy4 is live in Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, ima and the WeCom AI assistant. It also runs through OpenRouter, Hermes and OpenCode.

What's next: Tencent says it will publish the model weights and a system card soon. Open-weight means anyone can download the model and run it. A system card is a document that describes what the model can do and where it fails. Neither was out at launch.

A cheaper Hy4 Flash is also planned. Tencent describes it as the version for faster, lower-cost use.

The field is crowded. Mistral AI, Moonshot and Z.AI all offer open-weight models with 1 million token windows, so Hy4 is entering a race, not an empty track.

• Developers: you can try Hy4 through OpenRouter at a low price, and weights are promised soon.
• Office users: Tencent's apps may soon do multi-step jobs across programs for you.
• Everyone else: watch the hallucination numbers. Less made-up output is what makes these tools trustworthy.
• Skeptics: wait for independent tests before trusting any vendor's scores.