Stable Audio 3 · Medium
Generate audio and music from descriptions, with local model weights for custom workflows.
GenerationOpen-source code1.4B
- Maru integration
Maru workflow available
Source-workspace SDK integration. Requires an available inference service, Studio admission and an inference:run token; not a browser extension download.
- Start here
- Use Medium when a longer or more demanding audio task justifies it. For personal LoRAs, start from the matching base checkpoint, not the post-trained inference weights.
- Training data
- 1,278,902 recordings: 806,284 licensed from AudioSparx and 472,618 from Freesound; pretrained T5Gemma text conditioning.
- Training rights
- The card reports licensed stock audio and CC0, CC BY and CC Sampling+ Freesound recordings. T5Gemma has separate Gemma terms; a source attribution page is linked.
- License & output terms
- Stability AI Community License; inspect commercial conditions and component terms.
- Limitations
- A different release from Stable Audio Open 1.0. Hardware cost, prompt sensitivity and dataset bias remain relevant; terms are not an unrestricted open-source license.
- Technique
- Latent diffusion transformer
- Runtime
- Local inference with substantial model memory; consult the Medium configuration and upstream hardware guidance.
- Parameters
- 1.4B
Sources reviewed . Training and rights notes summarize provider disclosures. A model license does not clear every source recording or generated output. Parameter counts refer to the named variant or component, not total runtime memory.