Screenshots show the official pages on the capture date and may differ from the pages at publication. Select an image to enlarge it.
On November 4, 2026, Chinese tech giant Tencent released the official version of its AI model "Hy4." It follows the preview version released on August 28 and is the first in the Hy series (whose English name was changed from Hunyuan (混元) in December 2025) able to understand images. It also gains the ability to operate a computer while looking at the screen. Tencent says it plans to release the weights (the data that makes up the model) and a safety document (a system card) for the official version within the next few days.
Image understanding and computer use
The main new capabilities in Hy4 are as follows.
- Image understanding: reads charts and documents, and reasons across multiple images
- Computer use: operates command-line tools as well as on-screen buttons and menus
- Professional software: trained with particular emphasis on operating specialized software such as Blender for 3D creation and CAD for design
- Long tasks: handles long-running workflows that span multiple apps
The preview released in August had about 770B total parameters, with about 49B active per token, and the amount of text it could handle at once (its context) was 1M tokens. According to reports, the official version is the same size.
Performance
Some of the figures Tencent published are as follows. All are company-reported evaluations.
- A financial analysis evaluation: Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 (2.97) and Kimi K3.1 Pro (2.68)
- Rate of answers that do not match the facts (hallucinations): down from 12.95% in the preview to 7.2%
- A game development evaluation: up 19.1% over the preview, and up 9.8% on an evaluation of game mechanics
- An evaluation in WorkBuddy, the company's own AI agent product: an average score of 83.18 and a pass rate of 96.15%
Pricing and availability
API pricing is unchanged from the preview: per 1M tokens, about $0.90 for input, about $2.70 for output and about $0.045 for cached input. On Tencent Cloud's TokenHub, it is half price until December 31.
Hy4 is available in the following Tencent services.
- The AI assistant Yuanbao (元宝)
- The AI agent WorkBuddy and the coding tool CodeBuddy (free to use within certain limits from November 4 to 15)
- The AI assistant in the enterprise chat app WeCom, and the knowledge management tool ima
- Tencent Cloud's TokenHub
It can also be used from the third-party services OpenRouter, Hermes and OpenCode. Tencent plans to release a lower-cost "Hy4 flash" as well.
CoAI's take
China's major AI models are all moving in the same direction, toward agents that not only handle text but also understand images and operate computers. The fact that Hy4 chose the latest versions of DeepSeek and Kimi as comparison points shows how intense the competition among Chinese players is. The preview's weights were already released in August under the Apache 2.0 license as a text-only version, and the weights of the official version, which supports image understanding, are to follow.
On the other hand, the performance figures are Tencent's own evaluations, and external verification is still to come. The license for the official version's weights has not been announced yet either. For use from Japan, going through third-party services such as OpenRouter is the realistic option. If you use it for work, check how data is handled and the terms of service.
Sources
- Tencent launches Hy4, joining the ranks of the world's leading open-source AI models (Tencent)
- tencent/Hy4-preview (Hugging Face)
- Tencent launches Hy4 with vision and computer-use capabilities (TechNode)
- Tencent releases Hy4, a powerful frontier AI model introducing visual understanding and computer use (SiliconANGLE)