Stability AI unveils Stable Audio 3.0, an open-weight model family for artistic experimentation

Stability AI has unveiled Stable Audio 3.0, its next-generation audio model family designed to support artistic experimentation. All of these models were trained on fully licensed data, and 3 of the models are being released as open-weights, available for anyone to download and use.

The family consists of four models tailored to different use cases:

• Stable Audio 3.0 Small SFX: A sound effects generation model that can run on mobile and low-spec laptops
• Stable Audio 3.0 Small: A model capable of full on-device music composition
• Stable Audio 3.0 Medium: A model that delivers high-quality music structure and melody up to 6 minutes and 20 seconds long
• Stable Audio 3.0 Large: The flagship model for large-scale music platforms and professional environments requiring low-latency generation

Stable Audio 3.0 introduces a new semantic-acoustic autoencoder architecture to support longer and more flexible audio generation. Key features include:

• Variable-length generation: Audio can be generated for over 6 minutes with precise second-level control.
• On-device music composition: The Small model enables full music tracks—not just short samples—to be created offline.
• LoRa support: LoRa training is provided to fine-tune the model to match a user's own library.
• Audio inpainting: The models support editing features such as modifying specific sections of a track or extending existing songs (causal continuation).

Under the Stability AI Community License, users are free to distribute and commercially use generated outputs. However, companies with annual revenue of $1 million or more must obtain an Enterprise License to receive commercial protection and legal indemnification.