Tencent Open-Sources Hy-Nomas, an Interactive Video World Model That Reruns Google's Genie 4 Demos

Tencent's Hunyuan visual generation team has open-sourced Hy-Nomas, an interactive video world model that lets users steer a generated scene with free-form natural language while the world keeps evolving. The team published the technical report "Hy-Nomas: Towards Free-form Interactive World Generation" on arXiv (2610.36729), released the Hy-Nomas-Bench dataset for reproducible research, and put up a free, no-queue online demo at hy-nomas.hy.tencent.com.

Hy-Nomas is not a single video generator but a two-part system. The "director" is a Prompt Enhancer, a vision-language model fine-tuned from Alibaba's Qwen3.8-27B with supervised fine-tuning. It reads the user's current instruction together with the full interaction history and every clip generated so far, then converts that into a structured, more detailed generation brief and decides how long the next segment should run. The "camera" is an 8-billion-parameter video diffusion backbone initialized from Tencent's own open-source HunyuanVideo 1.5 and retrained for continuous generation, instruction following and camera control.

Coordination between the two is what separates Hy-Nomas from an ordinary text-to-video model. A history cache carries the state of the world forward, action-time constraints keep user commands synchronized with the timeline, and explicit camera controls let the operator move through a scene that persists between prompts. In one example described in the report, a user asks the system to turn a cat into an old-fashioned radio; the prompt enhancer has to work out where the cat currently sits in the frame, how the transformation should unfold, and what else in the scene changes, before handing the plan to the video backbone.

The most eye-catching claim is a direct comparison with Google's Genie 4, the interactive world model Google shipped in late October. Within roughly 48 hours of Genie 4's demos appearing, the Tencent team says it reproduced eight of the official sample clips with Hy-Nomas and published side-by-side videos. One architectural gap remains: Google's setup generates audio together with the picture in real time, while Hy-Nomas currently produces visuals only.

All numbers in the comparison come from the Tencent team itself and have not been independently verified. What is verifiable is the release itself: the arXiv report, the open code, the Hy-Nomas-Bench dataset and the live trial page are all public, which makes this one of the first interactive world generators that developers can inspect, benchmark and run rather than just watch in a demo reel.

The release lands in an increasingly crowded lane. Google's Genie 4 doubled interactive worlds to two minutes with native audio; NVIDIA, Odyssey, VAST and others have all shipped world models this quarter, and site coverage has tracked each step. Most of those remain closed or partially open. An open interactive world model with a published benchmark gives academic groups and smaller studios something they have not had: a baseline they can rebuild on.

The practical question is whether open weights can close the gap on responsiveness. Genie 4 streams audio and reacts within a live session; Hy-Nomas generates clips that respond to the next prompt. For research and creative pipelines, that difference matters less than access. For real-time applications, Tencent will need a follow-up that closes the audio loop.