World models are having a crowded week: while DeepMind's Genie 4 is still making waves, Tencent's Hunyuan team has posted the technical report for Hy-Nomas, its own interactive world generation model, on arXiv. A paper author put it bluntly in the Hugging Face comments: Hy-Nomas supports exactly the same features as Genie 4, but is fully open-sourced. The confidence comes from a complete artifact drop — weights, code, a benchmark dataset, and a lightweight adaptation workflow.
What it does
Hy-Nomas is a video world model built on an 8B diffusion backbone. It takes multi-turn, free-form text instructions or continuous camera trajectories, generates variable-length video segments, and keeps a cached history so the world stays consistent — objects introduced in one segment are still there in the next. The central design aligns structured video captions with action-aware causal attention: descriptive context stays accessible across segments, while action instructions are restricted to their intended temporal windows, preventing old actions from contaminating new frames. A vision-language Prompt Enhancer translates user requests and multimodal interaction histories into structured conditions for the next segment. Training follows a multi-stage curriculum — causal generation first, then text interaction and camera control, then training on model-generated histories, followed by diffusion distillation.
It runs on a single RTX 4090
More telling than "yet another world model" is the deployment bar. Hy-Nomas ships under Apache 2.0, with an official low-memory configuration for a single RTX 4090: DiT weights and KV cache stored in INT8 (computation stays in BF16), sliding-window history truncation, CPU offload, and FlexAttention plus CUTLASS FP8 decoding, combined with a distilled 4-8-step denoising loop. The ecosystem moved within a week: training and inference code, weights and an online demo landed October 21, a Diffusers example on October 23, WanGP support on October 25, SGLang and ComfyUI integrations on October 26, and the arXiv report on October 27. On the Google TPU side, a community developer measured real-time 720p generation: about 0.933 seconds of compute per second of video on a v6e-4 slice (4 chips), with no late frames during a 60-second run.
The caveats
The abstract claims Hy-Nomas outperforms existing world models in human evaluation — that is the team's own framing, with no independent benchmark to cross-check yet, and "exactly the same features as Genie 4" is likewise an author's claim. Even discounted, though, the completeness of the drop is unusual for this track: model weights, the Hy-Nomas-Bench benchmark, a LoRA fine-tuning path, and an annotation workflow for preparing training data from local videos — most of what reproducible research needs is already in the GitHub repository. While closed leaders show off demo reels, the open-source answer is "you can run it yourself." When do world models graduate from demos to production? Now anyone with a 4090 can go find out.