Stability AI released the Stable Audio 3.0 family on May 20, with four models unified and extended to 6 minutes 20 seconds: small SFX, small (459M), and medium (1.4B) are released as open-source weights, while large (2.7B) is API and enterprise self-hosted. The key change is not just the length doubling — it's putting "fully licensed data" on the product front page: licensing partnerships with Warner and Universal, downloadable weights, GitHub open source, avoiding the copyright-litigation reefs that Suno and Udio have been navigating.
The 459M small targets on-device real-time generation (≤2 minutes), the 1.4B medium holds the 6-minute long-structure line, and the 2.7B large is for multi-track production. Combined with the "chunked diffusion + temporal conditioning" continuation architecture, melody, harmony, and rhythm maintain consistency in the token-level latent space — this is the capability threshold that Suno v4 and Lyria 3 Pro have been working toward but never explicitly stated.
The commercial posture is more intriguing: Ethan Kaplan, formerly of Fender, was hired to lead the professional-musician product line, and large goes enterprise-licensed for customers with over $1M annual revenue. The "open source + B-end licensing fee" combination is the first sample that simultaneously puts "open" and "licensed" on the table.