Tencent open-sources Hy-Nomas world model: 18.5 FPS on 8×B200 with full history

Tencent has released the code and weights of Hy-Nomas, its world model controlled through text and camera. Everything is available under the Apache 2.0 license, which means the model can be used in commercial projects and fine-tuned on your own data. In measurements on the Reactor test bench with full history on 8×B200, it generates about 18.5 frames per second. There is a separate low-resource config for a single RTX 4090, but the authors do not state its speed.

The announcement came a few days before Genie 4, so the open release landed at almost the same time as the new version of a closed model in this class.

How Hy-Nomas builds a world frame by frame

The model does not generate a clip in one go. It builds the video in chunks and stores the context between them in a KV-cache. As a result, each new chunk builds on what has already been generated instead of starting from scratch: the scene stays coherent while the user moves the camera through the space.

Control works on two levels. Text sets what happens in the scene. Camera movement can be described in advance or changed between chunks, which means the user can steer the picture during generation. For a world model, this is a basic requirement: without a controllable viewpoint, you just get an ordinary video generator.

Hy-Nomas joins a growing set of open models with the same control scheme. MIRA, Solaris, and Odyssey-3 sit alongside it, and all four support text and camera input. The market now has real options to choose from, rather than just closed demos to watch.

License, LoRA, and ready-made adapters

For developers, openness matters most here. The code and weights are published under Apache 2.0, one of the most permissive licenses. Fine-tuning is possible through LoRA on your own videos. You do not need to retrain the whole model: it is enough to train a small adapter for the style or domain you need.

The authors included two ready-made adapters right away. The first tunes the output to a cinematic look, and the second is designed for robotics. This shows where the team is aiming: controllable video simulation for film and creative work, as well as for environments where robots are trained and their behavior is tested. How well the robotics adapter works in practice can only be checked through the technical report and your own tests. The original announcement contains no quality metrics.

The release comes with a demo, a technical report, code, weights, and a page with speed measurements. That means the performance figures can be double-checked in the repository itself.

What is known about speed and hardware

Precise wording matters here. The 18.5 FPS figure was obtained on the Reactor test bench: full history and 8×B200. This is a data-center-class configuration, and the result cannot be extrapolated to consumer hardware.

For a single RTX 4090, the repository has a separate config. It enables three restrictions:

• INT8 quantization;
• limited history, meaning the model sees fewer previous frames;
• offloading of models to RAM.

The speed of this mode is not stated. Since the history is cut and part of the weights is offloaded, you can expect losses in scene coherence and in speed. But that is an assumption, not a measurement. Anyone who wants to run Hy-Nomas on a home card will have to test for themselves whether an interactive frame rate is achievable.

Practical takeaway

Hy-Nomas will be useful to teams that research controllable video simulation and want to work with an open model rather than a closed API. The license does not restrict use, and LoRA lowers the cost of adapting the model to your own domain. Before adopting it, read the technical report and run the model on your own scenes. The announcement does not confirm real-world coherence on long camera trajectories or speed on a consumer card. The main measured result is 18.5 FPS on 8×B200, and not everyone has access to that configuration.