• Variable-length generation up to 6:20, at per-second granularity

• Full song composition on-device (first open model to do this)

• Audio inpainting: single-segment editing, multi-segment editing, and causal continuation (extending a track beyond its original endpoint) LoRA training support with published documentation

• Fully licensed training data

• Commercial use permitted under the Community License

The model family:

• Stable Audio 3.0 Small SFX — on-device sound effects, up to 2 min

• Stable Audio 3.0 Small — full music composition on-device, up to 2 min

• Stable Audio 3.0 Medium — higher musicality and longer tracks, up to 6:20