Stability AI Stable Audio 3.0

Stability AI has released Stable Audio 3.0, a new family of music and audio generation models that expands the company’s open-weight strategy into longer-form audio. The launch was posted by Stability AI on May 20, 2026, on its news site and product pages. The company said the family includes four models, with three available as open weights and one reserved for API and enterprise deployments. (stability.ai) Stability AI said the models were trained on fully licensed data and that users can commercialize outputs under its licensing terms. The release matters because Stability AI is pairing open distribution with a longer track-length target than its earlier audio systems. The company said Stable Audio 3.0 supports variable-length generation up to six minutes. That extends the timeline from Stable Audio 2.0, which Stability AI said in 2024 could generate songs up to three minutes long. (stability.ai) ### Which models are actually in the new release? Stable Audio 3.0 is described by Stability AI as a family of small, medium and large latent diffusion models for variable-length audio generation and editing, alongside SAME, a semantically aligned music autoencoder introduced with the release. The company’s product page says three of the models are open weights and free to download and build on. (stability.ai) Hugging Face listings show public model pages for stable-audio-3-small-music, stable-audio-3-medium and stable-audio-3-optimized, all updated around the launch window. Stability AI’s announcement said Stable Audio 3.0 Small and Medium are available on Hugging Face, while Stable Audio 3.0 Large is offered through the Stability AI API and for enterprise self-hosting. ### What changed from earlier Stable Audio releases? (stability.ai) May 20 marked Stability AI’s move from shorter generation toward full-length tracks. In its launch post, the company called out “variable-length generation up to six minutes” and “full song composition on portable devices” as key innovations in the new family. By comparison, Stability AI’s 2024 announcement for Stable Audio 2.0 said that model could generate songs up to three minutes long, with structured compositions including intro, development and outro. (stability.ai) The new release therefore doubles the stated maximum generation length from the prior version, based on the company’s own published specifications. ### What does “open weights” mean here in practice? (huggingface.co) Stability AI said the open-weight models are meant for experimentation and can be downloaded directly. On the Hugging Face model cards for the small and medium releases, the company said it is releasing weights for models that can run on consumer-grade hardware, together with training and inference pipelines. The same Hugging Face pages say the models were trained on licensed and Creative Commons data and can generate music and sounds in under two seconds on an H200 GPU and in a few seconds on a MacBook Pro M4. (stability.ai) Those performance claims come from Stability AI’s model cards and describe the company’s own test conditions rather than an independent benchmark. ### How does the licensing work for commercial use? Stability AI’s launch post said users “own your outputs” and can distribute and commercialize them under the Stability AI Community License, or under an enterprise license for organizations with more than $1 million in revenue. (stability.ai) The company’s license page says the community license covers research, non-commercial and commercial use for individuals or organizations generating under that revenue threshold. (stability.ai) Hugging Face model pages for the new releases list the license as “stable-audio-community.” That means the commercial-use claim attached to the launch is not a blanket public-domain grant; it is tied to Stability AI’s revenue-based licensing terms and separate enterprise arrangements for larger organizations. (huggingface.co) ### Where can users get it now? Stability AI’s product page says users can download the open weights from Hugging Face now and test the family through the company’s Stable Audio pages. The company said Stable Audio 3.0 Large is the model reserved for API access and enterprise self-hosting, while the small and medium releases are already public. May 22 Hugging Face listings still show the Stability AI audio repositories as recently updated, and Stability AI’s news page continues to point users to the May 20 launch post for the release details and download links. (stability.ai) (huggingface.co) (stability.ai) (huggingface.co)