HelixWorld license: Apache 2.0 code, weights and paper differ
HelixWorld's code is Apache 2.0, its weights have a separate license on Hugging Face, and the arXiv page lists CC BY-NC-SA 4.0. Read all three.
HelixWorld, Noiz AI's real-time audio-visual world model, has three licensing signals, and they are not the same. The GitHub README says the code is released under Apache 2.0 and that the checkpoint weights have a separate license published on Hugging Face. The arXiv listing for the paper shows CC BY-NC-SA 4.0. Do not assume that the Apache 2.0 code license covers the weights, and do not assume that the paper's license is the weights' license.
What each page says
The table lists what each page says. I read the GitHub README and the arXiv abstract page on 2026-10-13. I did not read the Hugging Face model card's license text, so I make no claim about what it allows.
| Where | What it says | Covers |
|---|---|---|
| GitHub README (NoizAI/HelixWorld) | Code under Apache 2.0 | The code in the repository |
| GitHub README | Checkpoint weights have a separate published license on Hugging Face | The preview v1 checkpoint |
| arXiv abstract page | License shown as CC BY-NC-SA 4.0 | The paper |
What this means for a commercial project
CC BY-NC-SA 4.0 is a non-commercial license. If it also governed the weights, commercial use would be out, but the README does not say that, so check the weights' own license first. If the weights license is non-commercial, a product that sells clips needs a different source for the video.
The README content I read does not mention a hosted or commercial API for HelixWorld. For a paid product, the practical alternative is a hosted video model with a price per second. On Sume, Gemini Omni Flash 1.1 and the other rows are billed per job through the same API, and you can read each row's current price from the catalog.
A short checklist
Before you use HelixWorld in anything you sell, write down the license text you read for each of the three artifacts, the date you read it, and the version of the checkpoint it applied to. The preview v1 checkpoint may be replaced by a later release with different terms.
- Read the Hugging Face model card's license, not only the repo.
- Record the checkpoint version.
- Keep the paper's license separate from the weights' license.
- Do not redistribute weights unless the weights license says so.
Sources
Related posts
More in Models
- Image-to-video API with sound: five-second cost by Sume row
Omni Flash 1.1 (always on), Wan 3.0 (audio), MiniMax H3 and H3 Max (native stereo) take a first frame and render sound. A 5 s clip is $0.19 to $0.94 on Sume.
- Is Qwen Image 2.1 Turbo on Sume? No: use these three Qwen rows
Sume's image catalog has no Qwen-Image-2.1 Turbo row. It lists Qwen Image, Qwen Image Max and Qwen Image Layered, and one call shows which.
- Long AI video clips on Sume: which rows pass 15 seconds
Only two Sume rows pass 15 seconds in one job: Seedance 2.5 (4-30 s) and Wan 3.0 (2-30 s). All other rows stop at 15 s or less. Read 2026-10-13.
- MAI-Voice-2.1 emotion control vs Sonic's six emotions on Sume
Microsoft lists granular emotion control for MAI-Voice-2.1. On Sume, Sonic 3 rows take six named emotions in generation_config.emotion; Eleven rows have none.
Written by Sume