THE ESSENTIALS IN 30 SECONDS

  • Hy-Nomas uses an 8B diffusion backbone and lets users steer each new video segment with free-form text and continuous camera control.
  • The paper says a 100-second, 24 fps video at 1280×704 peaks at 20.53 GiB of GPU memory with the single-RTX 4090 setup.
  • Weights, code and a 280-case benchmark are public; third-party write-ups list an Apache 2.0 license.

Tencent's Hunyuan team has published the technical report for Hy-Nomas, an interactive video world model that users steer turn by turn with written instructions and camera movements. The arXiv paper is dated October 27. The model itself arrived earlier and with little fanfare: according to OrcaRouter, the Hugging Face repository was created on October 21 and the weights and code were pushed on October 23, with no blog post or paper at the time.

How it works

Hy-Nomas is built on an 8-billion-parameter diffusion backbone. Rather than producing one clip from one prompt, it generates a series of variable-length video segments, each conditioned on the user's latest text and camera input plus a cache of what it has already generated.

The authors highlight two design choices. Structured video captions are paired with what they call action-aware causal attention, so that scene descriptions carry over from segment to segment while an instruction such as a movement or event only applies during its own time window. A vision-language "Prompt Enhancer" then turns loose user requests and the interaction history into structured conditions for the next segment. Training ends with a diffusion distillation stage, which cuts the number of denoising steps.

Running it at home

The default inference setup spreads work across eight GPUs, but the paper describes a single RTX 4090 path that stores weights and the attention cache in INT8, uses sliding-window attention and offloads to the CPU. In that mode, the authors report a peak of 20.53 GiB of GPU memory for a 100-second video made of seven segments at 1280×704 and 24 frames per second. That 24 fps is the playback rate; the report does not give generation speed, and AlphaSignal notes that the quality cost of the INT8 shortcuts is not quantified.

Claims and caveats

Tencent released the weights, code and Hy-Nomas-Bench, a set of 280 test cases, some requiring videos longer than a minute. OrcaRouter and AlphaSignal both list the model under Apache 2.0, though the Qwen2.5-VL text encoder it relies on keeps its own terms. In the paper, human raters compared Hy-Nomas pairwise with Genie 3, LingBot-World v2 and Happy Oyster, and the authors say their model came out ahead; those results have not yet been reproduced independently.