Tencent has released Hy4, a language model that can understand images and control software, expanding what the company’s model can do compared with its preview. The release comes about two months after that preview. Hy4 is available through Tencent Cloud, OpenRouter and Tencent apps such as the Yuanbao chatbot, but the model weights—the files needed to run or adapt it independently—will come later, on an unspecified date.
Tencent says Hy4 was trained on significantly more data than its predecessor. A company demo showed it searching a 2.8-gigapixel aerial image for a rooftop component about two meters long. It can also operate professional software, including Blender and CAD tools, and handle tasks that move between applications. Tencent says image understanding is useful for coding, too, and that it worked with its game studios to adapt Hy4 for Unreal, Unity and Godot.
How does it compare?
Tencent says Hy4 beat Deepseek V4.1 Flash, Kimi K3.1 Pro and GLM-5.3 Flash in a test of nearly 600 finance tasks judged by more than 200 finance professionals. That is a company-run evaluation. In Tencent’s own comparison across 12 benchmarks, Claude Opus 5.5 beat Hy4 on all 12, while GPT-6.1 Sol beat it on eight.
An independent Artificial Analysis evaluation puts Hy4 at 35 points on its Intelligence Index, above the 30-point median. That ties it with Grok 4.7 and places it second among Chinese lab models, behind Kimi K3.1 Pro at 46. But on Artificial Analysis’s factual-accuracy test without web search, Hy4 got about half the questions right from its training data. When it did not know an answer, it made one up 89 percent of the time.
What are the limits?
Tencent says its revamped web search reduced Hy4’s hallucination rate from nearly 16 percent to about 5 percent in internal tests. The model can also work with people on Word, Excel and PowerPoint files in real time. Those results do not remove a clear weakness: Tencent acknowledges that Hy4 trails leading commercial models on advanced math, long coding projects and complex image analysis. On the Hy-FrontierMath T5 test, Hy4 scored 13.0, compared with 97.5 for GPT-6.1 Sol and 91.3 for Claude Opus 5.5.
Hy4 costs about $0.83 per million input tokens and $2.50 per million output tokens, the same rates Tencent charged for the preview. Tencent plans a faster, cheaper Hy4 Flash version. It also says a safety report will arrive with the weights; until then, outside users cannot inspect or run those weights themselves.
Why it matters
People using Tencent’s services can try Hy4’s image analysis and software-control features, but they cannot yet download its weights to run or adapt it independently. Its price is set, but its performance is uneven: the independent test found a substantial weakness in answering from memory when it does not know the answer.
