Tencent cloned Google's Genie 4 and gave it away for free

A couple of days ago Google showed off Genie 4, a model that builds interactive worlds on the fly. Tencent's reply was blunt: on the evening of October 29 it posted a video recreating Genie 4's examples with its own model, Hy-Nomas. Same scenes, different AI.

Hy-Nomas is the same kind of world model. You type what happens and steer the camera like in a game, with WASD keys, and the model draws the world as you go. Video is built in chunks, and the model remembers what it already showed, so the scene holds together when you turn around. You can set the camera path in advance or change it between chunks, and drop in events with text.

The big difference is that Tencent gives it all away. Code and weights are on GitHub and Hugging Face under Apache 2.0, you can fine-tune it on your own videos with LoRA, and there are ready adapters for cinematic footage and robotics. There's a low-memory INT8 mode for a single RTX 4090, though its speed isn't listed. On serious hardware, eight Nvidia B200s, it hit about 18.5 frames per second.

The model itself came out on October 21, the technical report landed on the 27th, and now Tencent has put it head to head with Google. The online demo has no waitlist, so you can try it right now.

Google showed the future, and Tencent put it on GitHub.