Hy-Nomas
Tencent's Apache 2.0 interactive world model that continues a video world segment by segment from text and camera commands.
In short. Hy-Nomas is a Tencent model that builds a video world you can keep steering. You write what should happen next or give a camera path, and it makes the next few seconds of video, remembering what came before. It is Apache 2.0 and has a special setup for one RTX 4090, though the default setup needs eight GPUs.
Summary
Tencent published Hy-Nomas on Hugging Face on 21 October 2026, and its technical report appeared on arXiv (2610.36729) on 27 October. The report calls it a video world model with an 8B diffusion backbone for free-form interactive world generation. It builds a world one video segment at a time, and between segments you can add objects, change the style or scene, describe actions or move the camera, using text instructions and camera trajectories. Segments can have different lengths, and the card's example uses two 5-second segments. A persistent key-value cache carries the earlier segments forward so that later ones stay consistent with them.
The backbone is a dense multimodal diffusion transformer with 54 double-stream blocks and a hidden size of 2,048, with segment-wise causal attention, camera cross-attention and the language-model part of Qwen2.5-VL-7B as the text encoder. The released distilled model denoises in 8 steps and outputs 720 by 1280 pixels at 24 frames per second, and a fine-tuned PixelUMM decoder with 36 layers turns the latents into video. The repository also holds the non-distilled base model, a Cinematic LoRA, a robotics LoRA trained on a real-world robot-arm dataset, and Nomas-PE, a prompt-enhancer model that the web demo requires. Tencent released Hy-Nomas-Bench too, a set of 280 test cases for text interaction, camera navigation and spatial memory. The paper says human evaluation shows better results than existing world models, which is Tencent's own evaluation.
The standard setup uses 8-way sequence parallelism, so it needs eight GPUs in one group, and the card does not name the GPU model. A separate configuration targets a single GeForce RTX 4090 with an INT8 DiT of 9.1 GB, an INT8 cache, an FP8 decoder of 9.1 GB and CPU offload. It trades speed for memory, needs enough system RAM for the offloaded models and the growing history archive, and cannot load the LoRA adapters. The code expects Linux, CUDA 13.0 and PyTorch 2.10, and Tencent recommends its Docker image. The whole repository is about 196 GB because it keeps FP32 files and the extras. In the 30 days to 11 November 2026, Hugging Face counted about 90,000 downloads.
Good for
- Research on interactive, multi-turn world generation with text and camera control
- Previsualization and game or simulation experiments where a scene must stay consistent across turns
- Robotics simulation experiments through the robotics LoRA
- Trying a world model on a single RTX 4090 with the low-memory setup
Watch out for
- The default setup needs eight GPUs and the card does not say which, while the single-GPU route offloads to the CPU and adds latency
- Quality claims come from Tencent's own human evaluation and its own benchmark
- It needs Linux, CUDA 13.0 and PyTorch 2.10, and the attention kernels are built from source unless you use Tencent's Docker image
- The web demo depends on the separate Nomas-PE prompt enhancer
- The card documents text and camera-path input only, not image or video prompts
- Apache 2.0 covers Hy-Nomas itself, while third-party parts such as the Qwen text encoder keep their own licenses
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Hy-Nomas. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Distilled DiT (FP32 file, 8-GPU standard setup) FP32 | 8.55B | 34 GB | 39 GB | Workstation |
| INT8 DiT with FP8 decoder (single-GPU low-memory setup) INT8 | 8.55B | 18 GB | 24 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 24 GB of GPU memory.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is Hy-Nomas?
- Hy-Nomas is a Tencent model that builds a video world you can keep steering. You write what should happen next or give a camera path, and it makes the next few seconds of video, remembering what came before. It is Apache 2.0 and has a special setup for one RTX 4090, though the default setup needs eight GPUs.
- Can I use Hy-Nomas commercially?
- Yes. Hy-Nomas is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Hy-Nomas open source?
- Hy-Nomas is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Hy-Nomas on my own computer?
- Yes, on a good graphics card. The smallest variant, INT8 DiT with FP8 decoder (single-GPU low-memory setup), needs about 24 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Hy-Nomas released?
- Hy-Nomas was released by Tencent on October 21, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Hy-Nomas model card (Hugging Face) huggingface.co
- Hy-Nomas: Towards Free-form Interactive World Generation (technical report, arXiv) arxiv.org
- Hy-Nomas code repository (GitHub) github.com
- Hy-Nomas online demo hy-nomas.hy.tencent.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Cosmos 3 NVIDIA | 3.86B – 64.6B | – | OpenMDW-1.1 (Permissive) | 11 GB Laptop or phone | |
| Wan 2.2 Alibaba Wan | 5B – 27B 14B active | – | Apache 2.0 (Permissive) | 24 GB Gaming GPU | |
| LTX-2.5 LTX (Lightricks) | 22B | – | LTX-2.x Community (Conditional) | 16 GB Laptop or phone | |
| MiniMax H3 MiniMax | 33.1B | – | MiniMax H3 Community (Conditional) | 166 GB Big-memory desktop |