On May 20, 2026, Stability AI launched the Stable Audio 3.0 family — four models in one release: Small SFX, Small, Medium, and Large. The company positions it as a family of open-weight models “built for artistic experimentation.” Medium and Large generate full compositions of up to 6 minutes 20 seconds while holding musical structure and melodic tone, more than double what Stable Audio 2.0 managed in 2024. Three of the four ship as open weights, and the entire family is trained on fully licensed data.
That combination lands squarely on the two sorest points in the music-generation market: output length and training-data legality. While Suno and Udio remain entangled in litigation with the recording industry, Stability walks around the battlefield using contracts it signed with Universal and Warner.
Four Models, From Phone to Data Center
- Small SFX (459M parameters): on-device sound effects for phones and consumer laptops
- Small (459M parameters): full music composition on-device, up to 2 minutes; Stability calls it the only model capable of full music composition on-device
- Medium (1.4B parameters): improved structure, melodic coherence, and phrasing, with tracks up to 6:20
- Large (2.7B parameters): the strongest musicality in the family, aimed at high-volume, latency-sensitive music platforms
The baseline comparison is brutal. Stable Audio Open, the open model from June 2024, topped out at 47 seconds; the smaller Open Small variant managed 11. The new Small pushes on-device capability to two full minutes. Variable-length generation works at per-second granularity — for sound design in video, games, and ads, round-number track lengths are the exception, not the rule.
Open Weights and the Community License Line
Small SFX, Small, and Medium are downloadable on Hugging Face under the Community License: you own your outputs and can distribute and commercialize them freely. Organizations with more than $1 million in annual revenue, however, must buy an enterprise license, which adds indemnification and optional white-glove fine-tuning assistance. Large stays closed — available only through the Stability API and enterprise self-hosting, with ComfyUI and other partner platforms to follow.
Licensed Data as the Moat
The sharpest line in the announcement is the jab at competitors: Stability claims rival open music models either restrict commercial use or risk having been trained on unlicensed music. The confidence comes from two 2025 contracts — a strategic alliance with Universal Music Group in October and a next-generation tooling partnership with Warner Music Group in November. While competitors explain in court where their training data came from, licensing itself becomes the differentiation. Stability’s market thesis is blunt: “Artist-centric AI will only win if the product experience on a licensed platform is better.”
Technical Highlights and Hiring Signals
- A novel semantic-acoustic autoencoder architecture, with a research paper published alongside
- LoRA fine-tuning support with newly released documentation
- Inpainting for single-segment and multi-segment editing, plus causal continuation — extending a track past its endpoint
The hiring news carries its own signal. Ethan Kaplan, formerly chief digital officer at Universal Audio and Fender, joins to lead the professional music offering, with a musician-facing product suite in development and a waitlist open. The whole industry is recruiting music-industry veterans: Suno brought on ex-Merlin CEO Jeremy Sirota as chief commercial officer, and ElevenLabs hired Derek Cournoyer from Kobalt to lead music strategy.
What It Means for Developers and Creators
Three practical effects. First, on-device music generation finally has a usable open-weight option: two minutes of offline composition can live inside a phone app, and offline sound effects, game assets, and interactive installations are immediate beneficiaries. Second, the $1M revenue threshold draws a clean line — individuals and small teams use it free, companies pay for indemnification — which is becoming the standard posture for monetizing open weights. Third, with Google’s Lyria 3 Pro (March) and ElevenLabs’ music app (April) already in the market, Stability takes the third slot with “open weights plus licensed data.” The music-generation race is shifting from who can generate the longest track to who can commercialize legally.
Sources
- Meet Stable Audio 3.0 — Stability AI
- Stability AI releases a new audio model that can create 6-minute songs — TechCrunch
- Warner Music Group and Stability AI join forces — Stability AI
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.