Can You Use YuE2 Commercially? The Licence, Explained

Quick answer: No. YuE2's model weights are CC BY-NC 4.0, which prohibits commercial use of the weights and of music generated with them, and requires attribution. The inference code is separately Apache-2.0 and does permit commercial use. Confusing the two is the most common mistake made about this model.

YuE2 is downloadable, free and genuinely at the frontier of open music generation. None of that makes it commercially usable, and the gap between those two facts has caught out enough teams to be worth setting out carefully.

Four licences, not one

A single yes-or-no answer keeps going wrong because YuE2 is not under a single licence:

ComponentLicenceCommercial useModel weightsCC BY-NC 4.0NoFirst-party code, agent skill, docsApache-2.0YesOobleck / stable-audio-tools (in the VAE)MITYesSnakeBeta / BigVGAN (in the VAE)MITYes

The two MIT components sit inside modeling_vae.py and retain their own terms — Oobleck is copyright Stability AI, SnakeBeta copyright NVIDIA. The archived YuE-v1 branch keeps its original licence separately again.

Code and weights are not the same licence

You may read, modify, fork and commercially deploy YuE2's inference code under Apache-2.0. What you may not do commercially is use the weights — the trained model itself — or anything they produce.

Permissive code wrapped around restricted weights is the standard shape of an open model release. It is also easy to misread as a permissive release overall, particularly when a repository's licence badge reports only the code licence.

What CC BY-NC 4.0 actually requires

Two operative conditions, and a third thing people assume is present but is not:

• BY — attribution. Credit the authors, indicate changes, link the licence. This applies even to purely personal use.
• NC — non-commercial. No use primarily intended for or directed toward commercial advantage or monetary compensation. The wording is broad deliberately.
• No ShareAlike. You are not obliged to release derivatives under the same licence. This is the one respect in which it is more permissive than expected.

The phrase doing the work is “primarily intended for or directed toward commercial advantage”. It is deliberately wider than “you sold it”, which is why the grey areas below are genuinely grey rather than lazily hedged.

Can I sell music made with YuE2?

Not under the published terms. M·A·P apply the restriction to output, not only to the model files. All of the following fall within it:

• Releasing a track commercially on any streaming platform.
• Licensing music for an advert, film, game or trailer.
• Monetising a video that uses YuE2-generated audio.
• Selling it as a stock or production-library asset.
• Using it as background music in paid client work.

What counts as commercial use

Some uses the licence addresses directly. Others it does not name:

UseWithin the NC restriction?Selling a track on a streaming platformRestrictedMonetised video using the audioRestrictedPaid client workRestrictedShipping the weights in a paid appRestrictedFree portfolio piece that wins you paid workNot enumerated in the licenceInternal demo at a for-profit companyNot enumerated in the licenceNon-monetised personal videoNot restricted; attribution requiredAcademic research and benchmarkingNot restricted; attribution required

The rows marked “not enumerated” are cases the licence text does not name either way. The operative wording is ‘primarily intended for or directed toward commercial advantage or monetary compensation’, and M·A·P have not published further guidance on how they read it.

Can I put YuE2 in a paid product?

Not under CC BY-NC 4.0. Distributing NC-licensed weights inside a commercial application falls within the NC restriction. This is the term that most often rules the model out for software teams.

Worth separating: shipping the weights inside your application, and providing a workflow that loads weights your user obtains themselves, are different arrangements. Open tooling commonly relies on the second. Neither changes the terms attached to the output.

Why open model weights carry NC licences

Worth understanding, because you will meet the pattern repeatedly. Research groups release weights to establish priority, invite scrutiny and enable reproduction — all of which need the model in public hands. An NC clause lets them do that while preserving the option to license commercially later, and while avoiding handing a competitor a free product.

It also means “open” in a model release routinely describes availability rather than permission. Downloadable is not the same as usable, and the distinction is invisible until someone opens the licence file.

By the Open Source Initiative's definition, CC BY-NC is not an open-source licence at all, because it discriminates against a field of endeavour. “Open-weights” is the accurate term for YuE2, and it is not a synonym for open source.

Is a commercial licence available?

None is published in the repository or on the Hugging Face model card. For NC-licensed models the usual route is to approach the authors directly, but at the time of writing no paid commercial option has been announced. No published option is not the same as no option.

What the licence permits

• Personal projects and non-commercial art, with attribution.
• Research, evaluation and benchmarking against other models.
• Learning the architecture, and building on the Apache-2.0 inference code.
• Non-monetised demos and educational material.
• Writing about it — coverage and analysis are unrestricted.

If you need commercially usable music AI

The route depends on what you actually require.

• Check each model's weights licence individually. Several open music models use permissive licences, but licences change between releases and a permissive predecessor guarantees nothing about its successor. Verify the version you intend to run.
• Read commercial providers' output terms. Paid services generally grant commercial rights on paid tiers, often not on free ones. The tier you test on may not be the tier that permits release.
• Approach M·A·P directly. NC licensors frequently do grant commercial terms on request.
• Keep YuE2 for non-commercial work. Prototyping, learning and personal projects remain entirely legitimate, with attribution.

How to check any model's licence

Three habits prevent this problem generally:

• Read the weights licence separately from the code licence, because they routinely differ and the repository badge usually shows only one.
• Check whether restrictions extend to output, since many licences that permit running a model restrict what you do with what it makes.
• Check third-party components inside the model, which carry their own terms regardless of the wrapper. YuE2 ships a THIRD_PARTY_NOTICES.md for exactly this.

Does running YuE2 on Promptus change any of this?

No, and it is worth being explicit because the question comes up whenever a model is available through a product.

YuE2 is available in Promptus as a CosyFlow, runnable locally on your own GPU or on Promptus cloud compute. Neither path alters CC BY-NC 4.0. A licence attaches to the weights and, as M·A·P intend it, to what those weights produce — not to the interface you used or whose hardware executed the model. Generating a track in a polished workflow does not make it commercially usable, and no tool can grant you rights its upstream licence withholds.

If your work is non-commercial, YuE2 through Promptus is a considerably easier route than a manual install. If you need to sell the result, the model itself is the blocker, and the fix is a different model rather than a different way of running this one. ACE-Step, also available in Promptus, is Apache-2.0 and permits commercial use; it scores lower on WildSongBench, and that is the actual trade-off in front of you.

A plain-language summary of the terms M·A·P publish for YuE2. The licence itself is the authority: read CC BY-NC 4.0 in full.

Sources: M·A·P YuE repository, its docs/benchmarks.md and THIRD_PARTY_NOTICES.md, and the YuE2-3B model card. Figures checked 15 September 2026 against the WildSongBench run published 12 September 2026.