Summary: The Stable Audio 3.0 family landed on May 20, 2026, and Stability AI did not stop at one model: the release spans Small SFX, Small, Medium, and Large. Cross $1 million in annual revenue, though, and an enterprise license becomes mandatory, bundling indemnification plus optional white-glove fine-tuning help. Open weights cover three of the four, and every model in the family was trained on fully licensed data.
The Stable Audio 3.0 family landed on May 20, 2026, and Stability AI did not stop at one model: the release spans Small SFX, Small, Medium, and Large. The company bills the set as open-weight models “built for artistic experimentation.” Medium and Large can carry a complete piece for as long as 6 minutes 20 seconds without losing musical structure or melodic tone — more than twice the reach Stable Audio 2.0 had in 2024. Open weights cover three of the four, and every model in the family was trained on fully licensed data.
Music generation has two open wounds right now — tracks that end too soon, and training data nobody can vouch for — and this release hits both. The recording industry’s litigation against Suno and Udio is still unresolved; Stability, holding signed deals with Universal and Warner, simply walks around that fight.
One Family, Four Sizes: Phone to Data Center
- Small SFX (459M parameters): sound-effect generation on-device, aimed at phones and consumer laptops
- Small (459M parameters): composes complete tracks on-device, up to 2 minutes; by Stability’s account, no other model can compose full music on-device
- Medium (1.4B parameters): stronger structure, steadier melodic coherence, and better phrasing, with 6:20 as the track ceiling
- Large (2.7B parameters): the family’s most musical model, built for music platforms with heavy traffic and tight latency budgets
Set the new models against their ancestors and the jump is stark. Stable Audio Open, released in June 2024, stopped at 47 seconds; Open Small managed just 11. The new Small reaches two full minutes entirely on-device. Length is adjustable with per-second granularity — and anyone doing sound design for video, games, or ads knows round-number track lengths are the exception in that line of work, not the rule.
Where the Weights Are Open — and Where the License Draws a Line
Hugging Face already hosts Small SFX, Small, and Medium, all under the Community License: outputs belong to you, and distributing or selling them is fair game. Cross $1 million in annual revenue, though, and an enterprise license becomes mandatory, bundling indemnification plus optional white-glove fine-tuning help. Large never gets its weights released; access runs through the Stability API and enterprise self-hosting, with ComfyUI and other partner platforms to follow.
The Moat Is Built Out of Licenses
The announcement’s sharpest edge points at rivals: in Stability’s framing, competing open music models either fence off commercial use or gamble that their training music was licensed. What backs that up is a pair of 2025 deals — October’s strategic alliance with Universal Music Group and November’s next-generation tooling partnership with Warner Music Group. Rivals have to account for their training data in court; Stability lets the licensing itself do the differentiating. Its market thesis is stated without hedging: “Artist-centric AI will only win if the product experience on a licensed platform is better.”
Fresh Capabilities, Fresh Faces
- A brand-new semantic-acoustic autoencoder architecture, released together with its research paper
- LoRA fine-tuning support, backed by freshly published documentation
- Inpainting that handles single-segment and multi-segment edits, plus causal continuation, which carries a track past its endpoint
The org chart carries its own signal. Ethan Kaplan — previously chief digital officer at Universal Audio and Fender — is joining to run the professional music offering, and a product suite aimed at musicians is in development with a waitlist open. Music-industry veterans are in demand across the board: Suno hired Jeremy Sirota, Merlin’s former CEO, as chief commercial officer, while ElevenLabs pulled Derek Cournoyer over from Kobalt to lead music strategy.
The Practical Read for Developers and Creators
Three takeaways. One, on-device music generation finally has an open-weight option worth using: two minutes of offline composition can now fit inside a phone app, and offline sound effects, game assets, and interactive installations benefit immediately. Two, the $1M revenue threshold splits the market cleanly — individuals and small teams keep paying nothing, companies pay when they want indemnification — a posture that looks set to become the default for making money from open weights. Three, Google’s Lyria 3 Pro (March) and ElevenLabs’ music app (April) got there first, and Stability claims the third slot with “open weights plus licensed data.” The contest is changing shape: less about who generates the longest track, more about who can commercialize legally.