AI Potluck
Back to Gap Map Model components / Image, video, 3D & music generation

MusicGen

Meta
restricted / Overall score: 2.8

MusicGen is Meta's transformer model that generates instrumental music from a text description, and in its melody variants from a text description guided by a reference melody. It comes in 300M, 1.5B and 3.3B sizes with mono and stereo builds, plus style-conditioned and stem-separated variants, and runs through Meta's AudioCraft library.

AudioCraft, the library that runs and trains MusicGen, is a separate entry.

Openness

2 high confidence
2.0
weights
open(every MusicGen checkpoint downloads from Hugging Face without a gate)
data
described(trained on 20K hours of licensed music from Meta, Shutterstock and Pond5
code
open(the MusicGen training pipeline is published in AudioCraft)
license
CC-BY-NC-4.0(the weights

The MusicGen weights are licensed for non-commercial use only, even though the code that trains and runs them is MIT. Meta publishes the training pipeline, and the licensed training music is described but not provided.

Adoption

4 high confidence
4.0

Adoption is measured as Hugging Face downloads across the MusicGen checkpoints, nearly all of them for musicgen-medium.

Capability

2 medium confidence
2.0

MusicGen places last on the instrumental music arena, a step below Stable Audio and far below the closed services. It is still one of only two downloadable model families ranked there.

Verified 2026-09-26