Tencent Hy-Nomas: an open interactive video world model you steer as it renders

Tencent Hunyuan released Hy-Nomas on Oct 23. It's an open interactive video world model. A normal clip model renders a whole shot from one prompt. Hy-Nomas works autoregressively instead, one segment at a time. Between segments you can give it a free-form text instruction or a camera trajectory, and a long-lived KV cache keeps the world consistent from one turn to the next. The weights are Apache-2.0 on Hugging Face. Hugging Face curators @HuggingPapers and @AdinaYakup picked it up the same day.

What the README lists: distilled inference in 8 denoising steps, 720×1280 output at 24 fps, a Qwen text encoder, a segment-autoregressive DiT and a PixelUMM decoder. It ships two LoRAs, Cinematic and dual-arm Robotics, plus a prompt enhancer called Nomas-PE. A low-memory recipe targets a single RTX 4090: INT8 DiT and KV cache, FP8 decoder, a sliding history window, CPU offload and FlexAttention. There's a Docker image, and it needs CUDA 13. A team member puts it at 8B parameters (https://x.com/XiyaoWang10/status/2113729052142154235). The technical report is still listed as 'coming soon'.

Why it's on ReelStack: previs and long camera moves. Being able to say 'pan left, now walk through the door' and get one continuous take is closer to directing than re-prompting separate clips. The documented LoRA recipe (rank 256, Muon, 2,000 steps on the non-distilled checkpoint, then merged into the distilled one) also means a studio could train its own look. Be aware that the example packing plan targets 128 GPUs.

Our take: 24 fps is the output frame rate, not proof of real-time generation. How fast it runs on one 4090 is unverified. Try the hosted demo before setting up CUDA 13. Spotted, not tested.