Tencent puts Hy-Nomas beside Genie 4. Remembering the world is harder.

A fresh comparison, not a fresh model

On 29 October, Tencent Hunyuan published a video recreating examples from Google’s Genie 4 with Hy-Nomas. It is a company-selected demonstration, not a controlled head-to-head benchmark. Hy-Nomas itself had already been released: the repository dates the code and demo launch to 21 October, and the technical report was submitted to arXiv on 27 October.

The offer is nevertheless concrete: a public project, downloadable model weights and inference code that researchers can inspect. That makes it possible to examine the machinery behind the examples, rather than judge the system only through a promotional edit.

Sources: Tencent Hunyuan / X [1] · Tencent Hunyuan / GitHub [2] · Hy-Nomas authors / arXiv [3]

The instruction arrives while the scene is unfolding

Hy-Nomas is an eight-billion-parameter diffusion model. It generates video segments from current instructions and cached history. A prompt-enhancement component uses previous visual context to help interpret the next request; action-aware attention limits an instruction to the relevant time window.

Sources: Hy-Nomas authors / arXiv [4]

The paper’s harder result

The paper compares spatial memory against Genie 3—not Genie 4. Across 60 scenarios involving an outward camera movement and a return, Hy-Nomas won 13.3% of comparisons, tied 56.7% and lost 30%. These are the authors’ human evaluations, not our tests.

The distinction is important because continuity is more than an attractive frame. A doorway, a landmark or an object’s position gives the viewer a map of the scene. If that map changes when the camera returns, an exploratory sequence loses its geography.

Sources: Hy-Nomas authors / arXiv [4]

A 4090 path exists. That is not an FPS guarantee.

The repository includes a low-memory route for a single RTX 4090. It combines quantisation, a sliding history window and CPU offloading. These reduce the working memory burden, but fitting a model on a GPU does not establish that it generates at interactive frame rates on that GPU.

Tencent also lists integrations with Diffusers and community interfaces. For a technical reader, the useful next step is to inspect the exact configuration and hardware notes. A demo running elsewhere is not a performance promise for a local workstation.

Sources: Tencent Hunyuan / GitHub [2]

Where cinema and games actually meet

For cinema, the appeal is being able to redirect a sequence instead of writing a new prompt for every isolated shot. For games, it is the possibility that a scene responds to an action rather than simply playing back. Both depend on what happens across time: whether an instruction takes effect, whether it persists for the right duration, and whether the surroundings remain coherent.

That does not make generated video an editable game scene. Pixels do not automatically provide meshes, collision rules or persistent game state. Hy-Nomas is interesting as an experiment in responsive visual sequences; treating it as a replacement for a game engine would ask the announcement to prove something it does not.

Sources: Tencent Hunyuan [5] · Tencent Hunyuan / GitHub [2]

The revealing shot is the return journey

A beautiful departure can be generated one frame at a time. A convincing return requires the system to preserve something about where it has been. Tencent has made that problem unusually inspectable by publishing code and reporting an unfavourable memory comparison alongside its stronger claims. The next useful demonstration is not merely another spectacular landscape. It is a place that is still recognisable when you come back.