Apache-2.0video-genRuns locally
Hy-Nomas: Specs, License & How to Run
Tencent Apache-2.0 interactive world/video DiT — text + camera, 8-step distill.
- License
- Apache-2.0
- Parameters
- Not yet verified
- HF likes
- 126 (2026-11-01)
- First seen
- 2026-10-21
TL;DR. Hy-Nomas is a video-gen model from Tencent Hunyuan. License: Apache-2.0. It is listed under Runway.
| Developer | Tencent Hunyuan |
|---|---|
| Type | video-gen |
| Architecture | Segment-wise autoregressive diffusion transformer (MMDiT). Card: 54 dual-stream blocks, hidden size 2,048, 16 heads (128 dim), Qwen2.5-VL-7B-Instruct language-model text encoder, camera cross-attention (extrinsics_attn), persistent KV cache, PixelUMM video decoder. Distilled inference: 8 denoising steps, 720×1280 at 24 FPS. Not LTX-2.5 and not Hy-Memory. |
| Modalities | video, text, camera |
| License | Apache-2.0 — LICENSE on tencent/Hy-Nomas is Apache-2.0 with a Tencent 2026 preamble (cardData.license apache-2.0, opened 2026-11-01). GitHub Tencent-Hunyuan/Hy-Nomas license API is NOASSERTION; we copy the LICENSE file. Third-party components keep their original terms. Not legal advice. |
| Hardware | Not yet verified — Card names a single NVIDIA GeForce RTX 4090 configuration using INT8 DiT weights, an INT8 KV cache, an FP8 decoder, and CPU offload. That device claim is vendor. HF usedStorage is 196,404,270,231 bytes (~183 GiB). Peak VRAM is not independently measured. Total parameter count is not stated as a single figure on the card sections we opened. (not independently tested) |
| Hugging Face likes | 126 (as of 2026-11-01) |
| Model card | https://huggingface.co/tencent/Hy-Nomas |
| First seen | 2026-10-21 |
Hy-Nomas is Tencent Hunyuan’s interactive world/video generator: free-form, multi-turn text plus camera control, one video segment at a time. Official hosts we opened on 2026-11-01: the ungated card tencent/Hy-Nomas (created 2026-10-21T02:17:37Z; last modified 2026-10-28T06:28:38Z; 126 likes / 53,401 downloads), the LICENSE file, github.com/Tencent-Hunyuan/Hy-Nomas (431 stars / 36 forks), the project site hy-nomas.hy.tencent.com, the report arXiv 2610.36729, and the bench set tencent/Hy-Nomas-Bench (280 test cases). API gated was false.
This identification page hangs off Runway alternatives. That mapping is editorial. We have no sourced cinematic bake-off against Runway, LTX-2.5, or Pika H3 Flex.
It is not Hy-Memory (memory writer) and not Hy-SimulMT (speech MT). Same Hunyuan org, different jobs.
What we verified (2026-11-01)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/tencent/Hy-Nomas (not gated) | HF API |
| Created / last modified | 2026-10-21T02:17:37Z / 2026-10-28T06:28:38Z | same |
| Likes / downloads | 126 / 53,401 | same |
| License tag + file | apache-2.0 + Apache-2.0 (Tencent preamble) | HF cardData + LICENSE |
| Library name | diffusion-single-file | HF API |
| usedStorage | 196,404,270,231 bytes (~183 GiB) | HF API |
| GitHub stars / forks | 431 / 36 | GitHub API |
| GitHub last push | 2026-10-30T03:38:53Z | same |
Card configuration (copied, not a benchmark): 8 DiT denoising steps; 720 × 1280 at 24 FPS; 54 dual-stream MMDiT blocks; hidden size 2,048; 16 heads; text encoder is the language-model component of Qwen/Qwen2.5-VL-7B-Instruct. A separate configuration names a single RTX 4090 path (INT8 DiT, INT8 KV cache, FP8 decoder, CPU offload). That device sentence is vendor.
The card does not print a single total-parameter figure in the sections we opened. We do not invent one.
How to run (from official docs, not tested)
The GitHub repo documents a web demo (Nomas-PE prompt enhancer is a separate service required by the demo) and offline inference from prepared prompt files. forkalt has not run either path. Pixel editors stay on PhotoCraft vs GIMP.
License
LICENSE is Apache-2.0 with a Tencent 2026 preamble and a third-party-component note. GitHub’s license API is NOASSERTION. Not legal advice.
What we did not invent
No quality ranking against Runway, LTX-2.5, Pika H3 Flex, or SenseNova 6.8 Code2Video. Peak VRAM and generation seconds/segment are not yet verified.
Related: LTX-2.5 — Pika H3 Flex — Pika H3 Flex vs SenseNova 6.8 Code2Video — Hy-Memory — Hy-SimulMT — Iris-3B (still images). Catalog: home. Corrections: contact.
Deploy options
- local: Hugging Face weights
- local: GitHub inference / demo
- hosted: Official project demo
FAQ
Are Hy-Nomas weights public?
Yes. huggingface.co/tencent/Hy-Nomas is ungated (created 2026-10-21T02:17:37Z; last modified 2026-10-28T06:28:38Z; 126 likes / 53,401 downloads in the 2026-11-01 API check; card license apache-2.0). Code is github.com/Tencent-Hunyuan/Hy-Nomas (431 stars / 36 forks; last push 2026-10-30). GitHub license API is NOASSERTION; the HF LICENSE file is Apache-2.0 with a Tencent preamble.
Is this the same as Hy-Memory or LTX-2.5?
No. Hy-Memory is a 30B-A3B memory-writer MoE. LTX-2.5 is Lightricks’ gated 22B audio-video DiT under the LTX-2.x Community License. Hy-Nomas is an interactive, segment-autoregressive world/video generator with camera control. See Runway alternatives.
Can I use it commercially?
The LICENSE file is Apache-2.0 with a Tencent preamble and a third-party-component note. Not legal advice.
Does it match Runway quality?
No sourced bake-off. Distilled 8-step / 720×1280 / 24 FPS figures are card configuration, not a ranking.
Sources & verification
- https://huggingface.co/tencent/Hy-Nomas
- https://huggingface.co/tencent/Hy-Nomas/raw/main/LICENSE
- https://github.com/Tencent-Hunyuan/Hy-Nomas
- https://hy-nomas.hy.tencent.com/
- https://arxiv.org/abs/2610.36729
- https://huggingface.co/datasets/tencent/Hy-Nomas-Bench
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.