Meta open-sourced MusicGen with a licensed dataset and an NC licence

AudioCraft released MusicGen's weights and code but split their licences, and named its training data source.

Historical event

August 2, 2023

First source published

June 8, 2023

Site publication

September 18, 2026

What happened

Meta's research team submitted the MusicGen paper, 'Simple and Controllable Music Generation', to arXiv on 8 June 2023, describing a single-stage transformer that generates music from text or a melody without the cascade of separate models earlier systems needed. The paper was later presented at NeurIPS 2023. On 2 August 2023, Meta followed with an announcement open-sourcing AudioCraft, the framework bundling MusicGen with a sound-effects model, AudioGen, and an improved EnCodec audio codec, releasing both code and trained model weights rather than describing a capability from behind a paper alone.

What the documents say

Meta's announcement states that MusicGen was trained on roughly 400,000 recordings totalling 20,000 hours of music 'that Meta owns or licensed specifically for this purpose' - a specific, checkable claim about provenance that many contemporaries did not make. The same announcement is candid about a limitation, stating its datasets 'lack diversity', noting in particular a skew toward Western-style music and English-only text and metadata pairs. The AudioCraft repository adds a detail the blog post does not foreground: the code is released under the MIT licence, while the model weights themselves carry a separate CC-BY-NC 4.0 licence, meaning non-commercial use only unless a different agreement is obtained.

Why it matters for makers

The distinction that matters is between open weights and an open licence to use them commercially, which are not the same grant, and AudioCraft is a clean example of the gap. A producer can download MusicGen's weights, inspect the code, and run it freely for research or personal, non-commercial work, but the CC-BY-NC 4.0 term on the weights means a track built from default MusicGen output is not cleared for commercial release without a separate arrangement, regardless of how permissive the surrounding code licence looks. Meta's own diversity caveat is a second mechanism worth naming: a model trained predominantly on Western, English-labelled music will condition less reliably on styles outside that distribution, a property of the training set, not a bug to be prompted around.

What to check before you use it

Before relying on any 'open' generative audio model, check the licence on the weights separately from the licence on the surrounding code, since the two are routinely split, as they are here. Check next whether the vendor's own documentation names a concrete dataset size and source, as Meta's does, or only asserts quality without describing provenance. This is an editorial checklist built from the pattern this release shows, not a term either document states as a general rule.

• Does the weights licence permit the specific commercial use intended, separately from the code licence?
• What language, genre or style distribution does the vendor disclose for its training set, and does the intended use fall inside or outside it?
• Is the dataset described as licensed, owned, or simply collected, and what does that distinction mean for the outputs?

AudioCraft's contribution to this record is documentary rather than musical: a major lab stated a concrete training-data size and licensing method in public, then split its code and weight licences in a way that rewards a maker who reads past the headline word 'open-source'.

Sources & reading trail

MusicGen paper describing the single-stage transformer architecture and its NeurIPS 2023 presentation.

Source published: 8 June 2023 · Retrieved: 16 September 2026

Meta's own announcement stating the training-data size and licensing basis, and the diversity limitation of the dataset.

Source published: 2 August 2023 · Retrieved: 16 September 2026

Confirms the split licence: MIT for code, CC-BY-NC 4.0 for model weights, and lists included models.

Source published: Not established · Retrieved: 16 September 2026

Papers, terms and official documents establish the record; the maker reading and the checks are Signal to Song editorial analysis. This retrospective draft does not imply the site published on the event date.

Continue reading

• MusicLM was published with its memorisation study attached
• Stable Audio switched from a licensed library to Creative Commons
• Open models show two ways to disclose training data
• Browse the complete the archive

Published September 18, 2026, not on the date of the event described.