Key Takeaways
- Hy-Nomas is an open-source interactive video world model from Tencent Hy, announced on October 23, 2026, and released under the Apache 2.0 license.
- Instead of generating one fixed clip, it produces video in segments that you can redirect with new text prompts and camera trajectories while the world keeps its memory.
- It is a 9B model optimized for inference on consumer GPUs, with reported 720p at 24 frames per second on a single RTX 4090.
- A lightweight LoRA adaptation lets you specialize the model for domains such as robotics; the official page is on hy-nomas.hy.tencent.com.
Quick Facts
| Type | Open-source interactive video world model |
|---|---|
| Devices | Consumer GPUs; official web page |
| Price | Free and open source (Apache 2.0) |
| Languages | Text prompts; check official site for languages |
| Highlight | Steer the scene mid-generation with text and camera |
| Learning curve | Moderate; prompt and camera control take practice |
| Category | AI & Models |
| Creator | Tencent Hy |
| Official site | hy-nomas.hy.tencent.com ↗ |
| Heat | 🔥 63.1K Heat |
| Last updated |
What Tencent Hy's Hy-Nomas Actually Is
Hy-Nomas is an interactive video world model built by Tencent Hy, the team behind the Hunyuan family of models. It was announced on October 23, 2026, and released as open source under the Apache 2.0 license, which means you can inspect, modify and build on it. The official page lives on hy-nomas.hy.tencent.com, and the release also points to code and model weights on GitHub and Hugging Face.
The core idea is different from a standard text-to-video generator. Rather than producing one fixed clip from one prompt, Hy-Nomas generates video in segments and lets you change the scene as it unfolds. You can feed in new text instructions and camera trajectories between segments, so the world you see can evolve instead of being locked in at the first frame.
That combination is why it is being described as a world model rather than a video tool. A world model tries to keep a consistent, navigable environment in its head, so objects, layout and motion stay coherent as you move through it. Hy-Nomas applies that idea to generated video, and the open license makes it unusually accessible for researchers and hobbyists alike.
- Open weights, permissive license
The model is released under Apache 2.0, so commercial and research use is broadly allowed. Check the official repository for the exact terms and any usage notes.
- 9B parameters
It is a 9-billion-parameter model, which is small enough to target consumer hardware rather than a data-center cluster.
- Built for interaction
Text prompts and camera trajectories are first-class controls, not afterthoughts. You direct the next segment rather than re-rolling a whole clip.
- Domain adaptation via LoRA
A lightweight LoRA adaptation lets you specialize the model for narrower fields such as robotics without retraining the whole network.
How Free-Form Interaction and World Memory Work
The workflow starts with an initial prompt that sets the scene. From there, Hy-Nomas generates a segment of video, and you can issue a new text instruction, a camera movement, or both to shape what happens next. Because each segment builds on the previous one, the model needs to remember what came before, which is where the world memory comes in.
World memory is the mechanism that keeps objects, lighting and spatial layout consistent across segments. If a red car is parked on the left side of the frame, memory is what stops it from vanishing or teleporting when you pan the camera or add a new instruction. This is the hardest part of any interactive video system, and it is the feature Tencent Hy highlights most.
Free-form interaction is the standout quirk: you can change the world after it has already started, not just at the first prompt. That means you can begin with a street scene, then ask for rain, then rotate the camera, then introduce a new object, all within one continuous generation. The result feels less like rendering a clip and more like steering a simulation.
- Segment-by-segment generation
Video is produced in chunks, and each new chunk accepts fresh instructions. This is what makes mid-generation changes possible.
- Camera trajectories as a control
You can specify how the virtual camera moves, which gives you cinematic control over pans, orbits and pushes rather than relying on text alone.
- Consistency is the point
World memory is designed to preserve identity and layout across segments. Test it with a distinctive object and see how long it survives.
Running Hy-Nomas: Hardware, Access and Practical Tips
Hy-Nomas is optimized for inference on consumer GPUs. The editor notes report that it can run 720p video at 24 frames per second on a single RTX 4090, which puts it within reach of a well-equipped desktop rather than a server farm. If you have a different GPU, expect performance to vary, and check the official repository for the current hardware guidance before assuming your card is supported.
You can access the project through the official page on hy-nomas.hy.tencent.com, and the release also points to GitHub and Hugging Face for code and weights. If you would rather not set up a local environment, the official site is the place to check whether an online demo is available and what its limits are. Availability of hosted demos can change, so verify on the official page.
When you experiment, start with a short clip and change only one instruction at a time. This makes it much easier to see how well the model retains objects and camera direction before you scale up to longer sequences. If you hit quality or speed problems, reduce resolution or segment length first, since those are usually the biggest levers on consumer hardware.
- Start small, then scale
Begin with a short clip and a single change per segment. Once consistency holds, increase length and complexity.
- One variable at a time
Change either the text prompt or the camera move, not both, when you are testing. You will learn the model's behavior much faster.
- Watch your VRAM
720p at 24 fps on an RTX 4090 is the reported target. Lower resolution or shorter segments are the usual fixes on smaller cards.
- Use LoRA for narrow domains
If you need a specific style or domain such as robotics, a lightweight LoRA adaptation is the intended path rather than full fine-tuning.
How Hy-Nomas Compares and Where It Falls Short
Compared with closed video generators, Hy-Nomas competes on openness and interactivity rather than raw cinematic polish. Many commercial tools produce a single high-fidelity clip from one prompt, but they do not let you redirect the scene mid-generation or run the model locally. Hy-Nomas trades some of that polish for control, transparency and the ability to run on your own hardware.
Compared with other open video models, its differentiators are the interactive segment loop, world memory and the LoRA adaptation path. Those features make it interesting for robotics and simulation-style research, where an agent needs an environment that responds to commands. The 9B size and consumer-GPU target also make it more approachable than models that assume multi-GPU setups.
The limitations are worth stating plainly. Interactive video is still hard, and consistency can drift over long sequences, especially when you make large changes. Performance depends heavily on your GPU, and some details such as exact supported prompt languages, hosted demo availability and licensing nuances are best confirmed on the official site. It is a strong fit for researchers, developers and curious tinkerers, less so for anyone who wants a one-click polished video.
- Interactivity vs. polish
Closed tools often win on visual fidelity, while Hy-Nomas wins on control and local execution. Pick based on whether you need to steer the scene.
- Open ecosystem advantage
Apache 2.0 licensing plus GitHub and Hugging Face distribution means you can inspect, adapt and self-host. That matters for research and product prototyping.
- Robotics and simulation angle
The LoRA adaptation for specialized domains such as robotics hints at use as a controllable environment rather than just a video toy.
- Known rough edges
Long-horizon consistency and hardware demands are the main caveats. Verify current limits and demo availability on the official page.
Step-by-Step Guide
Check the official page and repositories
Start at hy-nomas.hy.tencent.com to see the current release status, demo options and documentation. The announcement also points to GitHub and Hugging Face for code and weights, so confirm what is actually published before planning a setup.
Decide between hosted demo and local run
If an online demo is available on the official site, try it first to get a feel for the controls. If you want full control or offline use, plan a local install and check the repository's hardware requirements against your GPU.
Set up your environment
Follow the installation instructions in the official repository, including any dependencies and model downloads. If you are targeting consumer hardware, note the reported 720p at 24 fps on an RTX 4090 as a reference point and adjust expectations for your card.
Write an initial prompt and generate a short clip
Describe a clear, simple scene and generate a brief first segment. Keep it short so you can iterate quickly and observe how the model handles basic motion and layout before adding complexity.
Steer with one change at a time
Add a new text instruction or a camera trajectory for the next segment, but not both at once. Watch how well objects and camera direction persist, then gradually increase segment length or make bolder changes once consistency holds.
Frequently Asked Questions
Is Hy-Nomas free to use?
Yes. Hy-Nomas is released as open source under the Apache 2.0 license, so you can use, modify and build on it at no cost. That said, running it locally requires suitable hardware, and any hosted demo on the official site may have its own usage limits. Check hy-nomas.hy.tencent.com for the current terms.
What hardware do I need to run Hy-Nomas?
It is optimized for inference on consumer GPUs, and the reported target is 720p at 24 frames per second on a single RTX 4090. Other GPUs may work but with different performance, so check the official repository for current requirements. Lower resolution and shorter segments are the usual ways to fit smaller cards.
How is Hy-Nomas different from a normal text-to-video model?
A normal text-to-video model generates one fixed clip from one prompt. Hy-Nomas generates video in segments and lets you change the scene afterward with new text prompts and camera trajectories, while world memory keeps objects and layout consistent. That makes it interactive rather than one-shot.
What is the LoRA adaptation for?
A lightweight LoRA adaptation lets you specialize Hy-Nomas for narrower domains, with robotics given as an example. Instead of retraining the full 9B model, you train a small adapter that shifts its behavior. This is useful when you need a specific style or task without the cost of full fine-tuning.
Where can I try it or get the model?
The official page is on hy-nomas.hy.tencent.com, and the release also points to GitHub and Hugging Face for code and weights. Whether a hosted demo is available can change over time, so verify on the official site. If you prefer local use, the repository is the place to start.
