Stability AI opens the weights: Stable Audio 3.0 brings licensed music generation home
Stability AI released Stable Audio 3.0 on May 20, 2026: four music models trained on licensed data, three with open weights on Hugging Face. Producers own their outputs and can commercialize them freely under the community license, with tracks up to six minutes twenty and LoRA fine-tuning on your own material.
While everyone else sells AI music as a subscription, Stability AI just handed producers the weights. Stable Audio 3.0, released May 20, is a family of four models, Small SFX, Small, Medium and Large, trained on fully licensed data, and three of the four are downloadable from Hugging Face under the Stability Community License.
Open weights change the deal. You own your outputs and can commercialize them freely, with an enterprise license only kicking in above $1 million in annual revenue. The Small models, at 459 million parameters, are built for on-device generation. The Medium model stretches tracks to six minutes twenty. The Large 2.7 billion parameter model stays behind the API and enterprise self-hosting.
The technical upgrade is a new semantic-acoustic autoencoder enabling longer, variable-length generation, with single and multi-segment inpainting and causal continuation, meaning you can extend a piece past its original ending. Stability is publishing LoRA fine-tuning documentation, so artists can train the open models on their own sound, and everything plugs into ComfyUI for local setups.
Decrypt reports the release is backed by licensing partnerships with Universal Music Group and Warner Music Group. That combination, open weights plus major-label licensing, is the first of its kind, and for producers who want AI in the studio without their stems leaving the building, it is the most interesting release of the year so far.