Stability AI launches Stable Audio 3.0 with models generating up to six-minute tracks, three variants released as open weights

Stability AI unveiled Stable Audio 3.0, a family of four audio models. Three ship as open-weights models on Hugging Face: Stable Audio 3.0 Small SFX and Small (each 459 million parameters, generating up to two minutes), and Stable Audio 3.0 Medium (1.4 billion parameters, up to 6:20 minutes). The largest variant, Stable Audio 3.0 Large (2.7 billion parameters), is only available through the Stability AI API, partner fal.ai, or enterprise licensing.

Stable Audio 3.0 uses a new semantic-acoustic autoencoder architecture enabling variable-length audio generation with second-level control. The Small model is the only variant that enables full music composition on-device without sample limits—a shift from prior releases (Stable Audio Open Small topped out at eleven seconds; Stable Audio Open managed 47 seconds). The models include inpainting features for editing, modifying, or extending tracks.

All models were trained entirely on licensed data. Under the Stability AI Community License, generated audio can be used commercially for free up to one million dollars in annual revenue; organizations exceeding that threshold must contact Stability AI for enterprise licensing, which includes commercial coverage and legal indemnification.

Stability AI partners with Universal Music Group and Warner Music Group to back its licensing stance. The company emphasizes this positions it against competing open music models that either restrict commercial use or train on unlicensed data, citing legal risks that have affected competitors like Suno and Udio.