Meet Stable Audio 3.0, the model family built for artistic experimentation with open-weight models

Summary

Stability AI has introduced Stable Audio 3.0, a family of generative audio models trained exclusively on licensed data. The release includes four variants tailored to different deployment scenarios and performance requirements. The Small SFX and Small models support on-device operation for sound-effect generation and full music composition respectively, while the Medium variant extends track length to more than six minutes with improved structural coherence and melodic phrasing. The Large model targets higher-volume, low-latency use cases such as music platforms.

A central technical change is a semantic-acoustic autoencoder that supports variable-length output at per-second granularity. This replaces earlier fixed-length constraints and enables generation beyond six minutes in the Medium and Large models. The same architecture underpins new editing functions, including single- and multi-segment inpainting as well as causal continuation that extends an existing audio segment without regeneration of the full track.

Fine-tuning is addressed through published LoRa documentation and weights for the Small and Medium models, allowing users to adapt the base models to custom audio libraries. Licensing distinguishes between the Community License, under which generated outputs may be distributed and commercialized without restriction, and an Enterprise License that adds coverage and indemnification for organizations exceeding one million dollars in annual revenue. Open weights for three of the models are distributed via Hugging Face; the Large model is accessible through the Stability AI API.

Why it matters

Direct model release with actionable details on architecture, fine-tuning, deployment options and licensing. Product teams can download, experiment and integrate immediately; addresses workflow needs like on-device generation and customization.