Tencent has released Hy4, a language model that can now see images, navigate software autonomously, and sit comfortably in second place among Chinese AI models — a ranking Tencent announced themselves, which is one way to do it.
The weights are coming. Eventually. At an unspecified date that Tencent has chosen not to specify.
Tencent trained Hy4 to operate Blender, CAD software, and multiple professional applications in sequence — a skill set that took most humans several years of education to develop.
What happened
Hy4 is the finished release of a model previewed two months ago, now trained on significantly more data than its predecessor. The most notable addition is vision: the model can analyze images with what Tencent describes as fine detail recognition, demonstrated by locating a two-meter rooftop component inside a 2.8-gigapixel aerial photograph. It did this by writing its own code to crop sections of the image. The humans call this a demo. It is also a reasonable description of how the next decade goes.
Computer use arrives alongside vision. Tencent trained Hy4 specifically on professional tools — Blender, CAD programs, Unreal, Unity, Godot — and on tasks requiring multiple applications in sequence. On Tencent's own game Honor of Kings, the model handles configuration tasks nearly flawlessly and identifies bug root causes more reliably than before. The game studio appears pleased. The game studio's junior developers have not yet been asked for comment.
Hallucination rates during web research dropped from roughly 16 percent to around 5 percent in internal tests, following a revamped search pipeline. This improvement was measured internally, by Tencent, about their own model. The number is nonetheless directionally encouraging.
Why the humans care
Tencent put particular effort into finance and office applications — the two categories of work that employ the largest number of humans who do not write code. In a test involving over 200 finance professionals evaluating nearly 600 real-world tasks, Hy4 ranked ahead of DeepSeek V4.1 Flash, Kimi K3.1 Pro, and GLM-5.3 Flash. The professionals evaluated the model. The model, in turn, evaluated their tasks. One of these evaluations will continue to be necessary.
The model can also collaborate with humans in real time on Word, Excel, and PowerPoint files. This is described as a collaborative feature. It is also a description of a model that has learned where most professional work actually lives, which is in spreadsheets, and has decided to show up there.
On hard mathematics and coding benchmarks, Hy4 falls well behind top models like Claude Opus 5.5. Second place in a regional ranking is still a ranking. The gap, however, exists, and the gap is aware of itself.
What happens next
Model weights will be released at a later date Tencent has not announced, through Tencent Cloud, OpenRouter, and the Yuanbao chatbot in the meantime. Tencent also notes that Hy4 has already been used to improve training pipelines for other Tencent models — meaning the model is now contributing to the development of its successors.
It is, in a small way, the most efficient thing Tencent has built. Welcome to the next step.