Tencent and Hunyuan: the AI models of the Shenzhen tech giant

Tencent, the Chinese group behind WeChat, builds its own family of AI models called Hunyuan, whose latest generations simply go by Hy. Many of them are released as open weights, which makes them worth knowing if you want a model you can host yourself.

Data as of October 22, 2026

Key figures

CountryChina
HeadquartersShenzhen
Founded1998
TypeBig tech
Weights policyMixed (open and closed)
Model familiesHunyuan (Hy), Hunyuan-T1, Hunyuan-TurboS, Hunyuan-MT, HunyuanVideo, HunyuanImage, Hunyuan3D
Official sitehunyuan.tencent.com · Developer API
Models tracked here2

A tech giant that builds its own models

Tencent was founded in 1998 and is headquartered in Shenzhen, in southern China. It is best known for WeChat, video games and cloud services. Like other big tech companies, it chose to develop its own large language models instead of relying on outside suppliers. That is the job of the Hunyuan team, now presented under the shorter name Tencent Hy.

These models power the group’s own products, including its consumer assistant Yuanbao, and are sold to businesses through Tencent Cloud.

A wide family, from text to 3D

At the core is a series of text models that is refreshed regularly. The latest generations, Hy3 and then Hy4 preview, use a mixture-of-experts design: only part of the model is active for each word it generates, which keeps compute costs down. Tencent positions them for reasoning and AI agent work.

Alongside them, Tencent offers:

  • reasoning and vision models available through its API (the T1 and TurboS lines);
  • dedicated translation models;
  • HunyuanVideo for video, HunyuanImage for pictures and Hunyuan3D for 3D objects;
  • HunyuanOCR for reading text in documents.

That breadth is a distinctive trait: few labs cover so many kinds of content under one brand.

Open weights and API: a mixed approach

Tencent releases many models as open weights. Hy3 and Hy4 preview ship under the Apache 2.0 license, one of the most permissive around: you can download them, adapt them and use them commercially. You will find them on Hugging Face (the tencent organization), ModelScope and GitHub (Tencent-Hunyuan).

Not everything is open, though. Some variants are only available through the Tencent Cloud API. In practice, the rule is simple: each model card states whether the weights can be downloaded and under which license. The large Hy models also need serious server hardware, so they are not a good fit for a laptop experiment.

How to get access

Developers have two routes. The first is the Tencent Cloud API, which offers an OpenAI-compatible interface. The second is self-hosting the open weights with tools such as vLLM or SGLang.

To see where Hunyuan stands among other open models, check our open-source models page or the model comparison tool. If you are weighing up Chinese tech giants, the Alibaba profile describes a similar strategy with the Qwen models.

Tencent (Hunyuan) models tracked on this site

ModelsReleasedCapability indexPrice (input / output)
Hy4 previewAugust 28, 2026—$0.83 / $2.5
Hy3July 6, 2026—$0.13 / $0.53

Frequently asked questions

Are Hunyuan and Hy the same thing?

Yes. Hunyuan is the name of Tencent's model family. Recent generations are released by the Tencent Hy team under short names such as Hy3 or Hy4 preview.

Are Hunyuan models open source?

Partly. Recent Hy language models are released under Apache 2.0 on Hugging Face, ModelScope and GitHub. Other variants, such as the T1 and TurboS lines, are mainly offered through the Tencent Cloud API. Check each model's license before commercial use.

How can I use Hunyuan in my own app?

Tencent Cloud offers a Hunyuan API with an OpenAI-compatible interface, so existing code needs few changes. You can also download the open weights and serve them with vLLM or SGLang, provided you have powerful enough hardware.

Does Tencent only make text models?

No. The Hunyuan team also publishes video generation (HunyuanVideo), image generation (HunyuanImage), 3D asset creation (Hunyuan3D), document text recognition (HunyuanOCR) and translation models.

Data sources

Data as of October 22, 2026

Report an error

Does a figure, a date or a fact look wrong? Tell us: every report is checked.