Tencent has released Hy-Nomas as an open model for generating interactive virtual worlds, with local operation reported on an RTX 4090. Instead of producing only a video to watch, it lets users steer the camera and objects within the generated world.
Tencent AI Watch analysis
What happened
Hy-Nomas accepts text prompts and camera trajectories, and supports long-horizon generation with what Tencent calls world memory. The model is available under the Apache 2.0 licence. GIGAZINE reports that it can run locally on an RTX 4090 using techniques including int8 quantisation and CPU offloading.
Tencent’s announcement also describes lightweight LoRA adaptation for specialised areas such as robotics. A browser demo is available, though GIGAZINE says using it requires an account.
Why it matters
Interactive world models sit between video generation and environments people can actively explore. Camera control and object interaction make the result more than a moving picture, while the reported local setup gives technically equipped users a way to experiment without relying solely on a hosted demo.
The robotics angle is worth watching, but it is not proof that Hy-Nomas can control a real robot. A generated world can be a useful research environment; whether skills learned there transfer to physical machines is a separate question.
Our read
The interesting step is giving users room to act inside generated worlds, rather than just asking a model to render another clip. Open weights and a stated local route make this easier to investigate. The RTX 4090 requirement, even with memory-saving techniques, is still a high bar; this is not quite a casual laptop download.
What to watch
- Whether the model code and weights support the capabilities described in Tencent’s announcement.
- How well interactions and world memory hold up beyond the examples.
- Whether researchers demonstrate useful robotics work, and whether results transfer beyond simulation.
Discussion spark: If an AI-generated world is interactive and can be adapted for robotics research, what evidence would convince you that it is a useful simulation rather than an impressive-looking demo?
Sources and evidence
not affiliated with or endorsed by Tencent AI