Tencent has released Hy4, the newest generation of its large language model. For the first time the model can understand images and use a computer on its own, working in both command-line tools and ordinary app windows. Tencent kept the price of the earlier preview, about US$0.90 per million input tokens, and says it will publish the model weights and safety documents (system cards) in the coming days, which would make Hy4 an open model anyone can download.
The details
Tencent lists four main upgrades over Hy4 Preview, the test version that came before it:
- Vision: Hy4 can now take images as input. Tencent names detailed image processing, reading complex charts and solving maths problems shown in pictures as examples.
- Computer use: the model can operate command-line interfaces and graphical apps to carry out tasks. Tencent says it was given special training on professional software such as Blender and CAD tools.
- Longer tasks: better instruction following, better understanding of long documents and stronger performance on agent tasks that run over many steps.
- Fewer made-up answers: in Tencent's own tests the hallucination rate fell from 12.95% in Hy4 Preview to 7.2%, and errors caused by unsupported assumptions fell from 23.1% to 5.6%.
The company is also pitching the model at specific kinds of work: office tasks inside its WorkBuddy assistant (which connects to WeCom chats, email, documents, calendars and meetings), financial research, front-end and game development with engines such as Unreal Engine, Unity and Godot, and scientific research.
Who gets it and when
Hy4 is available now in a long list of Tencent products: Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, the Yuanbao assistant, WeCom's AI assistant and ima. Developers outside China can also reach the API through third-party platforms including OpenRouter, Hermes and OpenCode. Tencent says it is working on bringing the model to QQ, WeRead, Tencent News and the games Honor of Kings and Game of Peace.
| Item | Price or offer |
|---|---|
| Input tokens | about US$0.90 per million |
| Output tokens | about US$2.70 per million |
| Cached input tokens | about US$0.045 per million |
| Launch offer | 50% off on Tencent Cloud TokenHub until 31 December 2026 |
| WorkBuddy and CodeBuddy | free from 4 to 15 November |
A cheaper, faster version called Hy4 Flash is due soon, according to the announcement.
Numbers
- Finance test: in a blind review run by Tencent's Hy team with Tencent Financial Technology, 213 finance professionals rated answers to 579 real-world tasks. Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68). Tencent says this made Hy4 the best open-source model in that test.
- WorkBuddy: an average score of 83.18 and a 96.15% pass rate on WorkBuddy benchmarks, ranking first in nine of twelve scenarios in mainland China, by Tencent's count.
- Game development: 19.1% better than Hy4 Preview on a GameDev benchmark and 9.8% better on GameLogic.
All of these results come from Tencent or its partners; independent test results were not part of the announcement.
Background
Tencent says safety work focused on cybersecurity and biology. The model was trained on defensive security data, with material about exploiting systems left out, and the company's Xuanwu Lab and Keen Security Lab found no significant concerns about its attack abilities in the areas they checked. On biology, Tencent says the model has strong basic knowledge but is more limited than frontier models at designing open-ended experiments.
The Hy team also released two open research models built with Hy4: HyCell, for single-cell biology, and HyPDE, for physics fields.
What is still unknown
Tencent did not give the model's size or a firm date for the open weights, and it did not say which licence the weights will use.
Announced by Tencent in its official newsroom: read the announcement
