Inside a modern AI data center with rows of servers running large AI models like Tencent Hy4

What Is Tencent Hy4

China’s AI race just got a new headline act. On November 4, 2026, Tencent unveiled Hy4, the newest model in its Hunyuan family, and it is the first one in the series that can both see images and operate a computer on its own. If you have been watching open AI models catch up with the big closed ones, this release deserves your attention.

So what is Tencent Hy4, exactly? It is a large language model built on a mixture-of-experts design, released with an open licence, and aimed squarely at developers and businesses that want frontier-style abilities without renting them from a closed API forever. Here is what it does, how big it is, what it costs, and the one catch you should know about before planning around it.

What Is Tencent Hy4 in Simple Words

Tencent Hy4 is the latest flagship AI model from Tencent, the Chinese technology company behind WeChat. It belongs to the Hunyuan model family and replaces the earlier Hy4 Preview as the company’s most capable release.

The simplest way to understand it is this. Older Hunyuan models were text machines. You typed, they answered. Hy4 still does that, but it adds eyes and hands. It can look at an image and reason about what is in it, and it can work a computer through both a graphical interface and a command line, clicking, typing and moving through software the way a person would.

That combination matters because it pushes the model from “chatbot” territory into “agent” territory. A model that can see a screen and act on it can, in principle, complete whole tasks, not just describe them. If that idea is new to you, our guide to free AI agents explains how agent-style tools differ from ordinary chatbots.

What Can Tencent Hy4 Do

The headline features Tencent announced are these:

  • Image understanding. For the first time in the Hunyuan series, the model natively understands images, so it can describe, compare and reason over pictures as part of a conversation.
  • Computer use. Hy4 can operate software through a graphical user interface and through the command line, which is the skill agents need to do real work on a machine.
  • Long tasks. Tencent says the model was trained with longer agent workflows in mind, including work inside professional design applications, so it can keep going through multi-step jobs instead of answering one prompt at a time.
  • A very long memory for text. The context window reaches 1 million tokens, enough to hold entire codebases, long contracts or a shelf of research papers in a single conversation.

None of this makes Hy4 unique on its own. Rival labs have shipped similar tricks this year. What makes the release notable is the package: these abilities arrive together, in a model whose licence lets outsiders actually use it.

A professional using a desktop computer while an AI assistant works on screen, showing the computer use feature of Tencent Hy4

How Big Is Tencent Hy4

Hy4 is a mixture-of-experts model with about 770 billion total parameters. That number sounds enormous, and it is, but the trick of this design is that the model does not use all of it at once. For each query it activates roughly 49 billion parameters, the slice of the network it judges relevant to your request.

Think of it like a hospital. The building employs hundreds of specialists, but your appointment only brings three of them into the room. You get the benefit of the whole institution’s knowledge while paying, in computing terms, only for the experts who actually saw you.

This is the same broad approach used by several leading open models. It is also why releases like Moonshot’s Kimi AI draw so much interest: Chinese labs have become very good at building huge models that stay affordable to run.

An engineer with a tablet standing in a GPU server room used to train and run open AI models

Is Tencent Hy4 Really Open Source

Mostly, yes, with one honest asterisk. Tencent released Hy4 under the Apache 2.0 licence, which is about as permissive as software licences get. Companies can adopt the model, modify it and build commercial products on it without paying Tencent a royalty.

The asterisk is timing. At the launch announcement, the hosted API was live, but the downloadable model weights and the full safety report were still marked as coming in the following days. So on day one, “open” described the licence and the plan more than an actual download link. If you are reading this to decide whether your team can self-host Hy4 this week, check whether the weights have landed yet before you promise anyone a date. Anyone who has dealt with AI model deployment challenges knows that a model you cannot download yet is a plan, not a deployment.

What Does Tencent Hy4 Cost

If you use Tencent’s hosted version instead of running the model yourself, the announced API pricing is about $0.90 per million input tokens and $2.70 per million output tokens. Tencent Cloud is also offering a 50 percent discount on those rates through December 31, 2026, which makes early experiments unusually cheap.

For context, input tokens are roughly what you send the model and output tokens are what it sends back. A million tokens is a lot of text, somewhere around 750,000 words, so a small team can test Hy4 seriously for the price of a few coffees. Self-hosting, once the weights arrive, removes the per-token bill but replaces it with a serious GPU bill of your own, so the hosted route will suit most readers.

How Well Does Tencent Hy4 Perform

Tencent’s own testing gives an early picture, with the usual warning that a company’s numbers about its own model are a starting point, not a verdict. In a blind evaluation across 579 financial tasks, Hy4 scored 3.04, and Tencent says it improved on the iRaB benchmark by about 12 points over the Hy4 Preview model it replaces.

Independent results will matter more. The pattern with big launches is that third-party testers confirm some claims, deflate others and find strengths the maker never mentioned. Give it a few weeks. What is already clear is the direction of travel: a model with vision, computer control and a million-token memory, given away under an open licence, would have been unthinkable at this scale not long ago.

Who Should Care About Tencent Hy4

Three groups should pay attention. Developers get a powerful open model for agent projects, especially ones that need to see screens and act on them. Businesses get a cheaper route to AI features, with the option to self-host later for privacy or cost control. And everyone else in the AI industry gets a reminder that the open side of the race is not slowing down; it is speeding up, and much of the speed is coming from China.

If you just use AI tools rather than build them, Hy4 still matters to you, because models like this end up inside the apps on your phone. The assistant that books, summarises and organises for you next year may well be running on an open model released this month.

What is Tencent Hy4?

Tencent Hy4 is Tencent’s newest flagship AI model, launched on November 4, 2026 as part of the Hunyuan family. It is a mixture-of-experts language model that adds native image understanding and computer use to the series for the first time.

Is Tencent Hy4 open source?

Yes, it is released under the Apache 2.0 licence, so companies can use and modify it freely, including commercially. At launch, the hosted API was available first, with the downloadable weights and safety report promised in the following days.

How many parameters does Tencent Hy4 have?

Hy4 has about 770 billion total parameters, but because it uses a mixture-of-experts design it activates only around 49 billion of them for each query, which keeps running costs far lower than a model of its full size.

Can Tencent Hy4 use a computer by itself?

Yes. Computer use is one of its headline features. Hy4 can work through a graphical interface and a command line, clicking and typing through software to complete multi-step tasks, which makes it suitable for AI agent projects.

What does Tencent Hy4 cost to use?

Through Tencent’s hosted API, announced pricing is about $0.90 per million input tokens and $2.70 per million output tokens, with a 50 percent discount on Tencent Cloud through December 31, 2026. Self-hosting removes per-token fees but requires your own GPU hardware.

How much text can Tencent Hy4 remember at once?

Hy4 supports a context window of up to 1 million tokens, roughly 750,000 words. That is enough to analyse entire codebases, long legal documents or large sets of research papers inside a single conversation.

Leave a Reply

Your email address will not be published. Required fields are marked *