Open-Weight Stable Audio 3.0 Model Family Released for Artistic Experimentation

Stable Audio 3.0 is a music generation model family trained on fully licensed data, with 3 open-weight models released on Hugging Face. The goal is to expand community-driven experimentation in the audio domain, much like Stable Diffusion did for image generation.

• Stable Audio 3.0 Small SFX: A model for sound effects generation that runs on mobile and entry-level laptops.
• Stable Audio 3.0 Small: A model that supports full music composition on-device.
• Stable Audio 3.0 Medium: A model that delivers higher musical fidelity and generates tracks up to 6 minutes 20 seconds long.
• Stable Audio 3.0 Large: A high-performance model for music platforms and creative apps, optimized for low-latency, high-volume generation.

Licensing is split into a Community License and an Enterprise License. Under the Community License, users own the outputs and can distribute and commercially use them, while organizations with annual revenue over $1 million can obtain an Enterprise License for commercial coverage and legal indemnification.

Technically, a new semantic-acoustic autoencoder is at the core. This enables variable-length generation with second-level precision, allowing 3.0 Small to generate up to 2 minutes and 3.0 Medium and 3.0 Large to generate 6 minutes or more. The company states that 3.0 Small is the only model capable of full on-device music composition.

Additionally, the company has released LoRa training documentation for the first time, supporting fine-tuning based on its own library, and also offers audio inpainting features including single-segment editing, multi-segment editing, and causal continuation. 3.0 Large is available via the Stability AI API and self-hosted for enterprises, and new products for musicians are also in preparation based on collaborations with Universal Music Group and Warner Music Group.