Tencent Hy is the name of Tencent's family of large AI models. Until late 2025 it was called Hunyuan (混元), and many Chinese reports still use that name. Tencent says its Yuanbao assistant, its WorkBuddy agent and its ima knowledge app run on Hy models.
This explainer covers what Tencent Hy is, how the model generations fit together, which models you can download, and where Hy shows up in Tencent's apps. Facts are as of November 9, 2026, and each one links to the page it comes from.
Hunyuan and Hy: one family, two names
Tencent now uses "Tencent Hy" for the brand. Its August 5, 2026 release describes it as "Tencent Hy, formerly known as Tencent Hunyuan", and its July 6 release says "Tencent Hunyuan (now known as Tencent Hy)". A December 9, 2025 CNMO report on Tencent News said Tencent had renamed the brand from "混元 (Tencent Hunyuan)" to "Tencent HY", and that Tencent named "名字更短、传播更快" (shorter name, faster spread) as the core goal.
The Chinese team name is still in use. Tencent's Hugging Face cards for Hy3 and Hy4 preview carry the tag "hunyuan", the Hy4 preview card links to a ModelScope account named "Tencent-Hunyuan", and Chinese outlets often write 腾讯混元. So "Hunyuan", "HY" and "Hy" all point to the same model family.
Who makes Tencent Hy
The Tencent Hy team builds the models inside Tencent. Tencent says the team rebuilt its model infrastructure early in 2026. In the Hy3 release, Tencent writes that "from rebuilding its infrastructure in late January 2026, to the launch of Hy3 preview in April, and now to the latest progress of Hy3, Tencent Hy has completed its full model-development loop in less than six months."
Tencent also says it builds the models together with its own products. The Hy4 preview model card says the team "partnered with top experts inside Tencent", such as software engineers, game developers, finance analysts and security experts, and built training data "around the work they ship."
For how Tencent fits next to DeepSeek, Alibaba's Qwen, Moonshot's Kimi and the others, see our guide to China's AI labs.
The model generations
Tencent shipped both Hy3 and Hy4 in two steps: first a "preview", then an official release. Here is the record for 2026, from Tencent's own pages.
Hy3 preview (April 2026). Tencent's Hy3 release says the preview launched on April 23 "as the first model in the Hy3 generation."
Hy3 (July 2026). Tencent released the official Hy3 on July 6, 2026. Tencent describes it as "a hybrid fast-and-slow-thinking model built on a Mixture-of-Experts (MoE) architecture, with 295 billion total parameters and 21 billion active parameters, supporting a context length of up to 256K." A mixture-of-experts model only runs part of its network for each token, which is why the "active" number is much smaller than the total. Tencent made Hy3 available worldwide through WorkBuddy, Tencent Design Miora and Tencent Cloud TokenHub, according to its August 5 release.
Hy4 preview (August 2026). The Hy4 preview model card says the model "comprises 770B total parameters, of which 49B are activated per token" and lists a context length of 1M. The card says the attention design was "inspired by DeepSeek and GLM." The Hugging Face repository was created on August 27, 2026 (UTC).
Hy4 (November 2026). Tencent released the full Hy4 on November 4, 2026. Its release says Hy4 "introduces visual understanding capabilities for the first time" and adds computer use, meaning the model can operate software through "both command-line interfaces and graphical user interfaces." We covered that launch in detail in Tencent Hy4 Launches With Vision and Computer Use.
Hy4 Flash (announced). Tencent announced Hy4 Flash in October, as we reported in Tencent Hy4 and Hy4 Flash Announced. In its November 4 release, Tencent said Hy4 Flash "is scheduled for release soon" and gave no date, although a heading on the same page reads "Hy4 Flash Recently Released." Hy4 Flash was not out as of November 9, 2026.
Which Hy models you can download
Tencent publishes many Hy models with open weights, mostly on Hugging Face under the tencent account. As of November 9, 2026, the main language models there are:
- Hy3, with the license listed as Apache 2.0 on Hugging Face. Tencent's July release calls this a "commercially friendly Apache 2.0 license."
- Hy4 preview, also listed as Apache 2.0, plus an FP8 quantized version.
- Hy3 preview, with the license listed as "other" on Hugging Face, so check its license file before you use it.
The full Hy4 weights were not on Hugging Face as of November 9, 2026. In its November 4 release, Tencent said it would publish the model weights and system cards "in the coming days."
Tencent also posts smaller, specialized Hy models. Two recent ones on its Hugging Face account are Hy-Memory-30B-A3B (repository created October 15, 2026) and Hy-SimulMT-30B-A3B (created October 22, 2026), both listed as Apache 2.0. Hy-SimulMT is a real-time speech translation model, which we explained in Hy-SimulMT: Tencent's Open Real-Time Speech Interpreter.
If you plan to run one of these yourself, check the size first. The Hy4 preview card's deployment examples serve the FP8 version across 8 GPUs (a tensor-parallel size of 8) with the vLLM or SGLang server software, then call it through an OpenAI-compatible API.
Where Hy shows up in Tencent's products
Tencent puts Hy inside its own apps. According to Tencent's releases:
- Yuanbao is Tencent's AI assistant app. Tencent's Hy3 release says Yuanbao added an Agent function on Hy3 that can "generate files in formats such as PowerPoint, Word, Excel, PDF and HTML."
- WorkBuddy and CodeBuddy are Tencent products that the Hy4 preview card says Tencent co-designs Hy with. The Hy3 release calls WorkBuddy "one of the most widely used AI productivity agents in China", a claim Tencent makes about its own product. Our explainer on Tencent Buddy One and WorkBuddy covers how WorkBuddy links Tencent's office apps.
- ima is Tencent's knowledge-base app. Tencent says Hy3 improved ima's knowledge-base question answering.
- Marvis is described by Tencent as its "OS-level AI assistant."
- Weixin and games. Tencent says AI customer service assistants on Weixin Official Accounts and an AI assistant in the game Path of Exile: Advent on WeGame use Hy3.
For developers, Tencent sells API access through Tencent Cloud TokenHub. Its Hy4 release says the API "is being made available via third-party platforms such as OpenRouter, Hermes, and OpenCode." Prices and promotions change often, so check TokenHub or the third-party platform for the current rate.
How Tencent describes Hy's strengths
Tencent pitches Hy for work tasks. The Hy3 release lists "software development, office productivity, financial modeling, front-end design and game production" as areas of "particularly notable progress." The Hy4 preview card lists software engineering, office and analysis, game development and scientific research.
These are Tencent's claims. Tencent's benchmark results come from its own tests, and some use Tencent's own experts as raters. For example, the Hy4 preview card reports a blind test in which "163 internal experts rated model outputs on 203 engineering tasks", and says Hy4 preview "came out slightly ahead" of GLM 5.3 and Kimi K3. That test was run by Tencent and published with the model card in August 2026.
Tencent is also open about limits. The Hy4 preview card says it is "an early version of Hy4" and lists known issues, "among them, spending longer than necessary reasoning through complex tasks, and a tendency to over-verify its own work." In its coverage of the November 4 launch, Caixin reported that the team admits Hy4 still has a large gap with top closed-source models.
How Hy compares with other Chinese model families
Three traits stand out in Tencent's own descriptions of Hy:
- Built with products. Tencent says it co-designs Hy with its own products, and its Hy3 release says real-world use in WorkBuddy feeds into the models' continued evolution. Hy3 also runs in Weixin Official Accounts and in a game on WeGame. For two other Chinese approaches, see our explainers on ByteDance's Doubao and on DeepSeek.
- Open weights under Apache 2.0. Hy3 and Hy4 preview are listed under Apache 2.0 on Hugging Face. Not every Chinese lab does this: Baidu released ERNIE 5.5 Flash Base under a custom license.
- A preview, then a release. Tencent shipped Hy3 and Hy4 early as previews. Caixin reported that since Tencent rebuilt its AI infrastructure in February, Hunyuan has put out a major version about every two months, first as a preview and then as an optimized official release. The Hy4 preview card puts it this way: "we would rather ship early and hear what breaks."
Should you use it
If you use Yuanbao, WorkBuddy or ima, Tencent says you already use Hy. Outside China, Tencent's August 5 release names WorkBuddy, Tencent Design Miora and Tencent Cloud TokenHub as ways to access Hy3. If you want to run an open model on your own servers, Hy3 and Hy4 preview are options listed under Apache 2.0.
Before relying on any benchmark comparison, look for independent tests. Tencent's head-to-head numbers come from Tencent.
Sources
- Tencent: Hy3 official release (July 6, 2026)
- Tencent: Hy3 available globally (August 5, 2026)
- Tencent: Hy4 launch (November 4, 2026)
- Hugging Face: tencent/Hy4-preview model card
- Hugging Face: tencent/Hy3
- Hugging Face: tencent/Hy3-preview
- Hugging Face: tencent/Hy-Memory-30B-A3B
- Hugging Face: tencent/Hy-SimulMT-30B-A3B
- Hugging Face: tencent account
- CNMO via Tencent News: Hunyuan renamed Tencent HY (December 9, 2025)
- Caixin: Tencent releases Hunyuan Hy4 (November 4, 2026)
