Tencent Hy-Nomas: Open World Model, RTX 4090 Support and Memory Tradeoffs

Tencent has released Hy-Nomas, an open video world model that can continue a scene through successive text instructions and camera movements. Its official weight history records the initial upload on October 23, with code and downloadable assets now public.

An impressive clip is not enough to choose an interactive generation system. The useful question for a developer is what happens on the next instruction, and the one after that: does the experience stay coherent, responsive and affordable to run?

How Hy-Nomas generates a world

The model card describes segment-autoregressive video generation: each new segment can respond to text and camera control while using a persistent key-value cache from earlier segments.

A user can introduce an object, request an action or change the scene as generation proceeds. The release includes generation checkpoints, quantized assets, video decoders and LoRA adapters, rather than only a demonstration video.

For builders, the architectural distinction matters more than the label “world model.” An interactive sequence must respond to new instructions without losing what the viewer already saw. A convincing first clip is only the beginning of that test.

Useful checks would include returning the camera to a previous viewpoint, changing one object while preserving the rest of a scene, and continuing after several unrelated instructions. These are proposed evaluation cases, not results from testing this release.

Can Hy-Nomas run on an RTX 4090?

Tencent documents a single-RTX-4090 video-inference configuration. It uses INT8 model weights and cache, CPU offloading, and a bounded GPU history to reduce memory demand.

The web-demo deployment guide adds an important qualification: its prompt-enhancement service, Nomas-PE, remains required. A single GPU for video generation is not a promise that the entire interactive application, including prompt enhancement, fits on that card.

The guide also separates gateway health from worker readiness. A running web server does not establish that the model has loaded or can generate a segment. It recommends validating a text-generated segment, a camera continuation and the downloaded cumulative video.

That is a sensible distinction for anyone estimating deployment effort. Budget for model loading, host memory, data movement and service coordination, not only GPU capacity. Keep the first test small enough to diagnose failures before adding longer sessions or multiple users.

Why 24 FPS does not mean real-time generation

The model card explicitly identifies 24 FPS as the output video's playback rate. It is not a measured generation-throughput claim.

An interactive application also needs a usable delay between an instruction and the next visible result. Measure that delay separately from playback smoothness. Record it for both short and longer sessions: a fast first segment does not tell you how the system behaves after accumulating history.

Memory introduces another tradeoff. Restricting what remains on the GPU can make a workload feasible, but a deployment should still be evaluated for continuity and host-memory growth. Do not assume a bounded GPU cache means the whole session has a fixed memory cost.

License and practical starting point

The license file identifies Tencent's released code and weights as Apache-2.0, while retaining separate terms for third-party components.

The release provides enough material to investigate a real pipeline. It does not, by itself, establish physical accuracy, production reliability or suitability for training robots. Start with the interaction your application requires, then measure whether the model preserves the details that matter across it.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help teams explain emerging AI systems through concrete capabilities, deployment costs and the limits behind impressive demos.

Book a 30-minute call