AI Music Model Licensing: What You Can Actually Ship

“Open source” tells you almost nothing about commercial use. MusicGen’s code is MIT and its weights are non-commercial. Here is what every open music model is actually licensed under, verified against the model cards.

Last updated 2026-09-12

"Open source" tells you almost nothing about whether you can use an AI music model commercially. The single most expensive mistake in this field is reading MIT on a GitHub page and assuming it covers the model — for MusicGen, the most popular open music model there is, the code is MIT and the weights are non-commercial.

This page lays out what each model is actually licensed under, verified against the model cards rather than repeated from summaries.

Three separate questions

Almost every licensing mistake comes from collapsing these into one:

1. What licence is the code under? The GitHub repository — the inference scripts, the training pipeline. Usually permissive, usually irrelevant to your actual question.
2. What licence are the weights under? A separate file, often a separate host — the Hugging Face model card, or a MODEL_LICENSE in the repo. This is the one that governs whether you can use the model.
3. What can you do with the output? Not usually settled by either licence. Colour under the weights licence, shaped by training-data provenance and by your own jurisdiction's rules on what can be copyrighted.

AudioCraft is the canonical example: MIT code, CC-BY-NC 4.0 weights. The GitHub badge says MIT. You still cannot build a commercial product on MusicGen.

The matrix

Verified 12 September 2026 against each project's repository and Hugging Face model card.

The short version

ACE-Step being Apache-2.0 on both code and weights is genuinely unusual, and it is the main reason it appears in product work more often than its raw audio quality alone would justify.

Magenta RealTime is the other notable case: CC-BY-4.0 weights ask only for attribution, which is the most permissive licence on any generation model here. It generates short-form audio rather than songs, but if you need something commercially clean and real-time, it has no direct competition.

What "non-commercial" actually rules out

CC-BY-NC is more restrictive than people assume. It is not "don't sell the audio files":

• A paid app or service built on the model.
• A free app that is part of a commercial business — an ad-supported tool, a free tier feeding a paid one.
• Music in an advert, a commercial game, a monetised video, a client project.
• Internal use at a for-profit company. This one surprises people: the restriction is on commercial purpose, not on charging.
• Content monetised on a platform, which is most content on most platforms.

What it does allow: personal projects, academic research, teaching, and art you genuinely don't monetise.

BY also means attribution is required, which people skip even when they're within the licence. And the SA in AudioLDM 2 and Tango's CC BY-NC-SA adds share-alike — derivatives must carry the same licence, which can be viral in ways that matter if you fine-tune.

The Stability AI Community License

Stable Audio Open is the only model here on bespoke terms, and it sits between permissive and non-commercial. Broadly: research use is permitted, and commercial use is permitted for organisations below an annual revenue threshold, with separate terms required above it.

Both Stable Audio Open models are also gated on Hugging Face — you must accept the licence and share contact details before the weights download. That gate is the licence being presented to you, so a download implies acceptance.

OpenRAIL: permissive with strings

Riffusion is the one model here on a CreativeML OpenRAIL-M licence, inherited from the Stable Diffusion model it was built on — and it is a neat illustration of this page's whole point. The repository carries an MIT badge. The weights do not.

OpenRAIL is a third category, distinct from both permissive and non-commercial. It does permit commercial use, but attaches behavioural use restrictions — a list of things you may not use the model for — and requires that those restrictions flow down to anyone you pass the model or its derivatives to. So it is not a blocker for a commercial product the way CC-BY-NC is, but it is not MIT either: you inherit obligations you have to propagate.

In practice Riffusion is unmaintained and superseded, so this matters less for it specifically than as a pattern — any model fine-tuned from Stable Diffusion is likely to carry OpenRAIL terms regardless of what its own repository says.

Training data is a separate axis

The weights licence tells you what the authors permit. It does not tell you what risk you inherit from what the model was trained on, and the two vary independently:

Note the asymmetry: MusicGen has clean, licensed training data and non-commercial weights. Clean data does not buy you commercial rights, and a permissive licence does not prove clean data. Judge them separately.

Who owns the output?

The genuinely unsettled part, and where you should be most sceptical of confident answers — including the confident answers you'll find elsewhere.

• The weights licence colours the output in practice. Selling a track made with a CC-BY-NC model is a commercial use of that model, whatever you think about the audio itself.
• Purely AI-generated work may not be copyrightable in several jurisdictions — the US Copyright Office has held that human authorship is required. You may be free to use a track and unable to stop anyone else using it too.
• Human contribution matters. Arrangement, editing, performance and mixing on top of generated material strengthens any authorship claim.
• Style imitation is its own exposure. ACE-Step's own disclaimer flags this: output stylistically close to an existing artist can create problems no permissive licence resolves. Prompting for a named artist is the clearest way to create that risk.
• Disclosure is increasingly required. Several platforms now mandate labelling AI-generated music, independent of any licence.

Check it yourself in thirty seconds

Don't trust a table — including this one — for a decision that matters. Hugging Face exposes the licence in its API:

curl -s https://huggingface.co/api/models/facebook/musicgen-large \
  | python3 -c "import sys,json; d=json.load(sys.stdin); \
      print(d['cardData'].get('license'), '| gated:', d.get('gated'))"
# cc-by-nc-4.0 | gated: False

Then read the model card itself, and look for a MODEL_LICENSE or LICENSE_weights file in the repository — separate from LICENSE. Where those two differ, the weights file is the one that governs your use of the model.

A checklist before you ship

1. Find the weights licence, not the repository licence.
2. If it's CC-BY-NC or CC BY-NC-SA, stop — that model is out for commercial work.
3. If it's OpenRAIL, commercial use is allowed, but read the use restrictions and pass them downstream.
4. If it's bespoke (Stability's), read the current text and check where you fall against any threshold.
5. Check whether a fine-tune or LoRA you're using carries different terms from the base model.
6. Keep the attribution the licence requires.
7. Consider training-data provenance separately from the licence.
8. Avoid prompting for named artists.
9. Plan to disclose AI-generated content.
10. Re-check before each release. These terms move.

Next: the GPU and VRAM requirements matrix covers the other constraint on which of these you can actually use.

← AI music generation hub