MusicGen / AudioCraft
Meta's PyTorch library with inference and training code for MusicGen text-to-music, AudioGen and related audio models.
MusicGen, AudioGen, EnCodec, MAGNeT and JASCO in one PyTorch codebase.
AudioCraft is Meta's PyTorch library for audio generation research, and it is where MusicGen lives. It suits ML engineers and researchers who want to generate or train music and sound models, but the model weights are non-commercial.
What it is
The library holds inference and training code for MusicGen (text-to-music), AudioGen (text-to-sound), EnCodec, Multi Band Diffusion, MAGNeT, AudioSeal watermarking, MusicGen Style and JASCO. MusicGen ships ten pretrained models, from 300M parameters to 3.3B, some of which take a melody as extra input.
Who it's for
Suits researchers and developers prototyping music generation with their own GPU. It does not suit producers who want a ready-made app, and nobody planning commercial releases can use the released weights as published.
Key features
MusicGen text-to-music
Generate music from a text prompt, with melody-conditioned variants for text plus melody input.
Pretrained model sizes
Models range from musicgen-small at 300M parameters to large at 3.3B.
AudioGen sound effects
A text-to-sound model sits alongside MusicGen in the same library.
JASCO chord and drum control
A music model conditioned on chords, melodies and drum tracks.
AudioSeal watermarking
Includes audio watermarking code, with an option to train a watermark model.
Training pipelines
Training code is provided for EnCodec, MusicGen, Multi Band Diffusion and JASCO.
Up and running in 3 steps
Our honest verdict
What's great
- Code is MIT, and the library includes full training pipelines, so you can train on your own licensed data.
- The MusicGen documentation says the model was trained on 20K hours of licensed music, which matters when judging training-data provenance.
- Several tools in one codebase: music, sound effects, a neural codec and watermarking.
- A local Gradio demo and Jupyter notebook ship in the repo for quick experiments.
Where it falls short
- The released model weights are CC-BY-NC 4.0, so you cannot use them commercially as published.
- The MusicGen docs say medium-sized models (about 1.5B parameters) need a GPU with at least 16 GB of memory for inference.
- Install is pinned to Python 3.9 and PyTorch 2.1.0, with xformers and ffmpeg, which can clash with newer environments.
- There is no end-user interface or DAW plugin; output is audio files from Python or a demo app.
Pricing & license
The code is MIT-licensed and free; the weights are CC-BY-NC 4.0 and not licensed for commercial use. The README shows no paid plan or hosted service, so the only cost is the GPU hardware you run it on.
Open remote AI-training roles
876 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I use MusicGen output commercially?
The code is MIT, but the README says the model weights are CC-BY-NC 4.0 (non-commercial). Get legal advice before using generated audio in paid work.
How much GPU memory do I need?
The MusicGen docs state at least 16 GB for inference with the medium (~1.5B) models. Smaller models need less.
Can I try it without installing?
The MusicGen docs link a Hugging Face Space and a Colab notebook.
