Stability AI releases Stable Audio 3.0, a family of four open-weight generative audio models

Stability AI releases Stable Audio 3.0, a suite of four generative music models with three open-weight versions up to 1.4B parameters on Hugging Face.

Stable Audio 3.0 features four music generation models designed for distinct uses and deployments. Three of them, with parameters ranging from 459 million to 1.4 billion, are released as open-weight and downloadable on Hugging Face. The fourth, heavier at 2.7 billion parameters, remains reserved for professional use via the Stability AI API or enterprise-licensed hosting.

The suite allows for the production of sound effects, short tracks, or full songs, up to just over six minutes in variable length. The lightest models run on portable devices without a dedicated graphics card. The entire system relies on a novel architecture, built around a semantic-acoustic autoencoder that ensures longer and more flexible generation.

The outputs belong to the user, who can distribute and commercialize them under the Stability AI Community License up to $1 million in revenue. Training is based on an entirely licensed dataset. LoRA fine-tuning is supported and documented for the first time. In parallel, a waitlist opens early access to a future suite of tools for musicians.