AudioCraft weights are CC-BY-NC: which models that covers
AudioCraft's code is MIT but its README puts the model weights under CC-BY-NC 4.0, including MusicGen and AudioGen. What that means for ad and video audio.

AudioCraft's README splits the license in two: "The code in this repository is released under the MIT license", and "The models weights in this repository are released under the CC-BY-NC 4.0 license". The weights are what make sound, so for a paid ad the NC part is the one to read twice. The repository covers more than MusicGen, so the question "is MusicGen commercial?" is only part of the answer.
Which models live in the repository?
The README names the models it covers: MusicGen, AudioGen, EnCodec, Multi Band Diffusion, MAGNeT, AudioSeal, MusicGen Style and JASCO. The license statement is about the weights in the repository, so check each model's own card or file if you plan to use one for client work.
| Part | License stated in the README |
|---|---|
| Code in the repository | MIT |
| Model weights in the repository | CC-BY-NC 4.0 (LICENSE_weights) |
Why does that matter for video?
Most video audio ends up in a commercial setting: a product page, an ad, a client's channel. NC means non-commercial, so the weights suit research and personal use. The code license does not change what the weights allow. Teams often read the first line of a README, see MIT, and stop; the second line is the one that matters here.
What is the hosted alternative on Sume?
Music 1.0 says it runs on Google Lyria 3.5, and the Music Router routes sume/music-auto to Lyria 3.5 today, with lyria-3.5 and lyria-3-pro as explicit ids. A generation costs a fixed $0.125 on Music 1.0, and the router charges the fixed Music price per generation. Output is a Sume-hosted audio artifact, usually audio/mpeg.
That is not the same as an open weights model: you get no weights, no fine-tuning and no seed. You do get a URL that the Timeline accepts in soundtrack.url. What you may do commercially with a hosted engine's output is set by the terms that apply to you; read them as you would any license.
A short decision rule
Pick by where the output will be used:
- Research, a prototype or a personal project: AudioCraft weights are fine under CC-BY-NC.
- A client deliverable or monetized video: use a source whose license you can show, and keep the record.
- A bed inside a Sume render: a Music Router job, then Timeline
soundtrack.
What should you record?
Write down the repository commit or release you used, the license text on the date you read it and the model file name. For hosted jobs keep the job result. Licenses change, so a dated copy is worth more than a memory.
What about a research prototype that becomes a product?
This is the usual trap. A team tries MusicGen in a notebook, likes the result and wires it into a pipeline. By launch the weights are baked into a service that earns money, and the NC license is a problem nobody had budgeted for. Decide the license question on day one, while swapping the engine is still cheap.
If a swap is likely, keep the engine behind one interface. On Sume that interface is the Music Router request body: the same prompt and optional model, with one audio artifact coming back, so changing the engine id does not change your code.
Sources
Related posts
More in Models
- AuK: MIT speech model that edits audio, and what Sume covers
AuK is Tencent's 1.5B MIT-licensed speech model for generation and editing. What its card lists, what it omits, and which parts Sume's audio tools cover.
- 21:9 architecture photos with AI: which Sume models take the ratio
Nano Banana 2 and Pro, FLUX.2 pro and flex, Qwen and Recraft list 21:9 on Sume. GPT Image 2.5 gets 3840x1648 by custom size. Imagen, Ideogram, Seedream do not.
- Can you sell images from open-weights models? Licences compared
Open weights do not mean commercial use. Ideogram 4, Qwen-Image, FLUX.2 dev and LTX-2.5 differ on selling outputs. What each page says, and hosted rows.
- Cartesia voices speak up to 25 languages: voice plus language on Sume
Cartesia says 50+ library voices speak up to 25 languages natively. On Sume you send a voice id and a language code per job, and you test each pair.
Written by Sume