Stable Audio 3.0: open-weight music models you can run on your own machine, trained on stock and Creative Commons audio

Stability AI released Stable Audio 3.0 on 20 May 2026: four music and sound-effects models, three of them with downloadable weights. The pitch is a music model that runs on your own machine and was trained on licensed data. Both parts hold up. What "licensed" covers, and what the free licence lets you do, deserve a closer look than the launch post gives them.

What did Stability release?

The family has four models, according to Stability's announcement:

• Small SFX, for sound effects on phones and consumer laptops.
• Small, for music on-device, up to two minutes.
• Medium, for longer tracks, up to 6:20.
• Large, which runs past six minutes and is aimed at platforms that need high-volume, low-latency generation.

Small SFX, Small and Medium have open weights on Hugging Face. Large is only available through Stability's API or a paid enterprise self-hosted deployment. All of them generate to the second rather than in fixed lengths, and they support inpainting: you can regenerate one section or several, or continue a short recording. Stability has also published LoRA training documentation for the Small and Medium weights.

The jump in length is big. The previous open models topped out at 11 seconds (Stable Audio Open Small) and 47 seconds (Stable Audio Open), so this is the first open Stability model with enough length for a full song. Stability's research page says the models generate in under two seconds on an H200 GPU and in "less than a few seconds" on a MacBook Pro M4. It also says the training and inference pipeline has been released along with the Small and Medium weights.

TechCrunch reports that Stability is building a separate suite for professional musicians, led by Ethan Kaplan, formerly chief digital officer at Universal Audio and Fender, and has shared no features for it yet.

What does "fully licensed" actually mean here?

The Stable Audio product page calls the models "commercially safe" and trained on "fully licensed datasets". The announcement links Stability's deals with Universal Music Group and Warner Music Group in the same breath. It's easy to read that as "trained on major-label music". The model card says otherwise.

The Medium model card lists 1,278,902 training recordings from two sources:

• AudioSparx, a stock music and sound library: 806,284 licensed recordings.
• Freesound: 472,618 Creative Commons files (CC0, CC-BY and CC Sampling+), the same subset used for Stable Audio Open.

In the Freesound portion, audio tagged as music was sent to a content-detection company, and copyrighted material it identified was removed. The card doesn't name a label catalogue. Stability's own UMG announcement from 30 October 2025 describes building new professional tools together, and doesn't say UMG recordings train any Stable Audio model.

So the claim is true, but narrow. Licensed here means production-library and Creative Commons audio. That is a cleaner provenance story than most of the AI music market can tell. It also suggests (our inference, not something Stability states) that the base models will lean toward library-music idioms, and that your own fine-tuning will matter more than the base model for anything that sounds like a particular scene.

What the licence allows

The open weights come under the Stability AI Community License, and accessing them on Hugging Face is gated behind accepting its terms. Four terms matter for musicians:

• Free up to $1M a year. Use is free for people and organisations with annual revenue under US $1,000,000, counting affiliates. Above that you need an enterprise licence, and the community licence terminates from the date you cross the line. A new licence is at Stability's discretion. Commercial users must also register with Stability.
• You own the outputs, "to the extent permitted by applicable law." That qualifier matters. The US Copyright Office's January 2025 report says prompts alone don't make an output copyrightable; human expressive input does. A track you arrange, edit and play over is on firmer ground than a raw generation.
• No training rivals. You can't use the models or their outputs to create or improve another foundation generative model.
• Attribution if you ship it. If you distribute the model or a product built on it, such as a plugin or app, you have to show "Powered by Stability AI" and include the licence.

Enterprise licensees get legal indemnification. Community users don't, so if a dispute ever arises over a generated track, a solo producer carries it alone.

What it means for producers

The important change is where the model runs. With Small or Medium on a laptop and a LoRA trained on your own stems, your catalogue never leaves your drive. No platform sees it, logs it or adds it to anything. That is a real difference from hosted generators, where uploading reference audio usually means accepting the service's terms on what happens to it.

It is also more work. You handle the setup, the fine-tuning and the judgement calls about what counts as your own work. And the $1M ceiling sits awkwardly with how music income is booked: a label, publisher or management company that owns your fine-tuned model may cross it long before you do.

What we think

Stable Audio 3.0 matters less as a generator than as a precedent. A capable, local, fine-tunable music model with a published list of training sources moves the line for everyone else. A hosted service that won't say what it trained on, or that wants your uploads, now has to explain why, when an open alternative names its sources down to the file count.

The winners are working producers, sound designers and small developers. They get a tool they can bend to their own material without handing that material over. Library and stock-music composers are the awkward case. Their genre is closest to what the base model learned, and the model card doesn't say how, or whether, AudioSparx contributors were paid for this use. We would like Stability and AudioSparx to answer that. Labels gain leverage too: the UMG and Warner deals now look like the route to future models trained on commercial catalogue, on terms they set.

The direction is right: the model sits with the artist, works on the artist's material and runs inside the artist's workflow. Veena, a browser DAW with an AI co-producer, is one example of a tool where, its site says, everything you make stays yours. Our advice: read the Community License before you fine-tune anything you care about, keep your training stems and project files as proof of human authorship, and don't treat "commercially safe" as indemnity. On the free tier, it isn't.