How to run open music-generation models on your own computer

Every big AI music app runs in someone else's data centre. You type a prompt, maybe upload a hum or a stem, and it lives on their servers under their terms. Open-weight models flip that: you download the model once and generate on your own machine. Here is what that takes, model by model, and the licence trap to check before you put anything out. Model cards and docs were checked on 29 Oct 2026.

The three models worth knowing

MusicGen. The AudioCraft README says the code is MIT but the model weights are CC-BY-NC 4.0, and the MusicGen model card repeats that and lists research as the intended use. "NC" means non-commercial: fine for learning and sketching, not for a release or a client job. The card says it was trained on about 20,000 hours of licensed music from Meta's own collection, Shutterstock and Pond5, with vocals removed, so expect instrumentals. Sizes run from 300M parameters (small) to 3.3B (large), per the MusicGen docs.

Stable Audio Open 1.0. The model card caps output at 47 seconds of 44.1 kHz stereo and says it was trained on about 486,000 Creative Commons recordings from Freesound and the Free Music Archive. It sits under the Stability AI Community License, and the card sends commercial users to Stability's licence page; you also accept the licence on Hugging Face before you can download. Our Stable Audio 3.0 news story covers the newer open models from the same company, which run longer and use the same licence family, including its $1M annual revenue ceiling.

ACE-Step. The ACE-Step v1 3.5B card lists Apache 2.0, the most permissive licence of the three. It claims about 20 seconds to make 4 minutes of music on an A100, and warns that pieces longer than 5 minutes may lose structure. Its out-of-scope list rules out generating copyrighted content without permission and passing AI music off as human-made.

What hardware you need

• An NVIDIA GPU is the easy path. The MusicGen docs say medium models need at least 16 GB of GPU memory; smaller cards can still make short clips, or longer ones with the small model.
• Apple Silicon works, slowly. ACE-Step's own table shows about 2.27x real time on an M2 Max at 27 steps, against about 34x on an RTX 4090. Real time means a 1-minute track takes under a minute on the Mac, but only a couple of seconds on the 4090.
• CPU only is a last resort. The Stable Audio Open sample code falls back to CPU when there is no CUDA GPU; neither card gives a speed for that, and in our view you should treat it as "for testing the install".
• Disk. Plan for several gigabytes per model. Weights for a 1B to 3.5B model add up fast.

The steps, at a high level

1. Install Python and PyTorch. AudioCraft asks for Python 3.9 and PyTorch 2.1.0, plus ffmpeg, per its README. Use a separate virtual environment per model; their dependency pins clash.
2. Accept the licence where asked. Stable Audio Open is gated on Hugging Face, so log in and accept before downloading.
3. Download the weights. Either the project's install command or the Hugging Face library pulls them on first run.
4. Start the local interface. MusicGen ships a Gradio demo you launch with python -m demos.musicgen_app; it opens in your browser but runs on your machine. Leave off the --share flag if you don't want a public link.
5. Generate short, then long. Start at 8 to 10 seconds to check the install and speed, then raise the length.
6. Bring the audio into your DAW. Treat output as raw material. Our guide to fixing generated stems in your DAW covers the clean-up.

The privacy upside

Running locally means your prompts, melodies and reference audio never leave your computer, and no service can use them to train its next model. That matters more than it sounds: as our look at open weights versus closed apps shows, hosted app terms often license your inputs or outputs back to the company. It also means no subscription, no monthly download caps, and no service shutting down under you.

What we think

Local models are the best way to learn how AI music generation actually works, and the only way to keep unreleased ideas fully private. But read the licence as carefully as the hardware list. The most famous model here, MusicGen, cannot go on a commercial release at all, and "open" does not mean "free for anything". For a working producer today, ACE-Step's Apache 2.0 licence is the cleanest commercial fit, Stable Audio Open suits sound design and loops if you stay inside Stability's terms, and MusicGen is a sketchpad. None of these will match the polish of the big hosted apps on full vocal songs; what you get instead is control. For which hosted tools allow commercial use, see which AI music generators allow commercial use on a free plan.

We are not lawyers. Licences change, so re-check the model card on the day you download.