Tencent Open-Sources Hy-Nomas: A Video World With Memory at 720p on a Single RTX 4090

Ordinary video models are like a camera with a short memory: they generate a clip, and then every new chunk risks going its own way. A room you just saw looks different ten seconds later. Tencent has open-sourced Hy-Nomas, a "world model" that builds a virtual space segment by segment and takes into account what it has already shown.

For example, you walk down a generated street, type "let it rain" and turn the camera back around, and the street should stay the same, only wet. Memory is handled by a KV-cache (a cache of the transformer's internal states), output is 720p at 24 frames per second, and the model can run on a single RTX 4090 with 24 GB of memory. For developers, this is a chance to try such a model locally, without a cluster, but with caveats, which are covered below.

What Hy-Nomas can do

The model generates a world segment by segment. The user changes what happens with text commands, controls camera movement and continues generation with earlier scenes taken into account. To keep continuity between video chunks, Hy-Nomas stores context in a KV-cache, so the model does not have to "re-read" the entire history each time. The active history is limited, however, so the model's memory is not infinite.

The output is 720p video at 24 frames per second, and generation takes 8 diffusion steps. That is relatively few: the approach aims for practicality, not for maximum quality at any cost.

Under the hood and how to run it

The architecture is a segment-autoregressive MMDiT with 54 blocks. The Qwen2.5-VL-7B encoder handles text. DCAE-V compresses the video, and PixelUMM decodes it. Ready-made variants include LoRA adapters (small fine-tuned add-ons) named Cinematic and Robotics. The latter has also been adapted on video data from robotic manipulators, so the model can be tried not only for pretty scenes but also for simulating how a robot arm handles objects.

For those without a server farm, Tencent has made it possible to run the model on a single RTX 4090 with 24 GB of VRAM. This relies on INT8/FP8 quantization, a limit on active history and offloading part of the data to the CPU. But that does not mean real-time generation: Tencent has not announced an official speed on the RTX 4090, so don't expect game-like responsiveness. You will have to wait for each segment.

There is a second limitation: this is video. There is no 3D export, so you cannot load the scene into a game engine and walk around it freely. What you get is a sequence of frames that looks like a world.

License and ecosystem

The project is released under Apache 2.0, which also permits commercial use. However, some third-party components may have their own licenses, so it is worth checking the entire chain, including the encoder and decoders, before going to production.

The strength of an open release is already visible: third-party integrations have appeared for ComfyUI, SGLang, WanGP and Reactor, along with the first experimental applications. Yesterday Tencent added Prelulu to its list of third-party projects. Its authors show interactive mini-dramas with dialogue and alternative story branches. It is a useful pointer to where such models are heading: not toward a "dream game" but toward interactive video in which the viewer influences the plot.

What to do with it

If you build creative tools or game prototypes, Hy-Nomas is worth installing and trying: the demo, code and weights are already available, and the ComfyUI node plugs the model into a familiar pipeline. Roboticists may find the Robotics adapter interesting as a way to produce synthetic videos of manipulation. Product managers should consider whether users will tolerate the generation wait in scenarios such as interactive stories and training simulations.

Don't expect the release to be a ready replacement for a game engine: there is no 3D, no confirmed real-time performance, and memory is limited. But as an open foundation for experiments with worlds that remember their own history, it is one of the most accessible options available today.