Tencent has launched Hy4, adding vision and computer-use capabilities to its Hunyuan model. The practical shift is from answering prompts to handling visual material and tasks across command-line and graphical software, while Tencent says API prices remain unchanged from the preview.
Tencent AI Watch analysis
What happened
Hy4 can analyse images and charts, and Tencent says it can combine perception, reasoning and tool calls for visual tasks. It can also operate through both command-line and graphical interfaces. Tencent says it trained the model for longer tasks across applications and professional software including Blender and CAD.
The model is available through Tencent Cloud and OpenRouter, and in Tencent products including WorkBuddy, CodeBuddy, Yuanbao and ima, according to reports by BigGo and AIBase. Tencent’s release page, linked in BigGo’s account of the launch, is quoted as saying the model weights are planned for release; they are not described as available yet. A lower-cost Hy4 Flash version is also planned.
Tencent says Hy4 improves on its August preview and keeps the same API pricing. The company’s benchmark claims need to be read as its own results, not as an independent verdict on model rankings.
Why it matters
Vision plus computer use gives developers a model that can, in principle, interpret what is on screen and act through familiar software interfaces. That makes the release relevant to people building agents for office, research and development work, not only chatbot applications.
Availability through OpenRouter and existing Tencent products lowers the friction of trying it. The promised open weights could matter more for teams wanting to run or adapt the model themselves, but that benefit depends on the weights actually arriving.
Our read
This is a substantive release, not just a new version number: visual input and software operation widen the jobs Hy4 is intended to tackle. Try it against a real workflow before taking broad capability claims on trust, and keep the distinction between available APIs and future open weights firmly in view.
What to watch
- When Tencent publishes the model weights and system card.
- Whether Hy4 Flash arrives with the promised lower-cost, faster inference.
- How the model performs on real multi-step tasks across applications, beyond company-reported benchmarks.
Discussion spark: For an AI agent using your software, would you prioritise reliable computer control or the freedom to run and modify open model weights?
Sources and evidence
not affiliated with or endorsed by Tencent AI