Open-Source AI Music Models
The open-source AI music models you can run yourself — MusicGen, Stable Audio Open, ACE-Step, YuE, Riffusion, Magenta and RAVE — with their types, licences and hardware needs.
Models you can run yourself
Open models let you generate music without a subscription, keep your data private, and build them into your own tools — subject to each model's licence. Here are the ones worth knowing, all runnable in Python.
| Model | Type | Licence | Vocals |
|---|---|---|---|
| MusicGen | Text / melody → instrumental | Open, non-commercial weights | No |
| Stable Audio Open | Text → audio | Open (community licence) | No |
| Riffusion | Text → music | MIT | No |
| YuE2 | Lyrics → score → song | Apache-2.0 code, CC BY-NC weights | Yes |
| ACE-Step | Text → music | Open (Apache-2.0) | Yes |
| Magenta | Symbolic / MIDI | Open (Apache-2.0) | No |
| Magenta RealTime | Real-time audio | Open (CC-BY-4.0 weights) | No |
| RAVE | Neural synthesis | Open, non-commercial | No |
The short version
- MusicGen — the most established open instrumental model; great quality, but non-commercial weights. Tutorial →
- Stable Audio Open — short audio, loops and SFX; community licence. Tutorial →
- ACE-Step — fast, Apache-2.0 including the weights, does full songs; the most product-friendly licence here. Tutorial →
- YuE2 — the best-sounding open song model, and the only one that writes an editable score before rendering it. Its weights are CC BY-NC despite Apache-2.0 code, and it needs Linux with 24 GB of VRAM. Tutorial →
- Riffusion — the original open, spectrogram-based approach; now unmaintained, and the hosted app that shared its name has since become a different product entirely.
- Magenta and RAVE — symbolic/MIDI generation and neural timbre transfer, respectively. RAVE is CC BY-NC and built for live performance. RAVE tutorial →
- Magenta RealTime — the outlier: it generates in real time, streams on Apple Silicon, and its CC-BY-4.0 weights are the most permissive here. Short-form audio rather than songs.
Hardware. Most of these want a CUDA GPU with 8 GB+ of VRAM for comfortable use; the smallest MusicGen model will run (slowly) on CPU. Full-song models are the most demanding — YuE2 needs Linux and a 24 GB NVIDIA GPU with no quantised fallback, while ACE-Step runs in 8 GB and on Apple Silicon.
Full GPU & VRAM requirements for every model → · Full licensing matrix →
Not sure open is right for you? Compare with the web apps on the best AI music generators page.