Stability AI releases Stable Audio 3.0

Background

Text-to-audio systems take a written prompt and return sound: music, a sound effect, or a background ambience. A recurring constraint in this area is duration. A model that can produce a few seconds of coherent sound is not automatically able to produce a whole track, because longer outputs have to stay consistent with themselves over time — the same rhythm, instrumentation and texture carried forward rather than drifting. A second constraint is control: generation from a prompt alone gives the user no way to keep part of an existing recording and change only the rest.

What was released

Around 20 May 2026 Stability AI released Stable Audio 3.0 in four sizes: Small SFX, Small, Medium and Large ^[1]. The family is split deliberately: Small SFX is the smallest variant and is named for sound effects rather than full music, while the other three move up in capability. Three of the four models are open-weight, meaning their trained parameters are published for others to download and run. The largest model, Large, is not distributed that way; it is offered through an API and through enterprise self-hosting ^[1].

Key details

Medium and Large generate up to 6 minutes 20 seconds of audio ^[1]. That figure matters because it moves the output from a clip to something closer to a complete piece — long enough to hold a structure rather than a single idea.

The models also support audio inpainting and continuation ^[1]. Inpainting means filling in a designated gap inside an existing recording: the user marks the region to be replaced and the model generates new audio that has to fit what comes before and after it. Continuation means extending a recording past its current end, with the model picking up from the existing material instead of starting fresh. Both are editing operations rather than pure generation, and both depend on the model conditioning on audio that already exists.

On training data, Stability said all the models were trained on fully licensed data ^[1]. The accompanying paper describes the training data as licensed and Creative Commons data ^[2].

The accompanying paper

The release was documented in a paper submitted to arXiv on 18 May 2026 by Zach Evans, Julian D. Parker, Matthew Rice, CJ Carr, Zack Zukowski, Josiah Taylor and Jordi Pons ^[2]. The submission came shortly before the public release.

Why it matters

The release is best read as a packaging decision as much as a modelling one. Putting out a size ladder — a small sound-effect model, a small general model, a medium and a large — lets users pick a point on the cost and quality curve instead of taking a single hosted system. Releasing three of the four as open weights extends that choice to people who want to run the models themselves, while keeping the largest behind an API and enterprise self-hosting.

The 6 minute 20 second ceiling and the editing features point at the same goal: audio generation used inside a production workflow, where a track has to be long enough to be usable and where existing recordings have to be repairable rather than replaceable. Stating the licensing position of the training data addresses a question that has followed generative audio closely, since the output of such systems is often intended for commercial release.