Stability AI releases Stable Audio 3.0 with models capable of generating six-minute compositions

Stability AI announced Stable Audio 3.0, a new family of audio generation models consisting of four variants with parameter counts ranging from 459 million to 2.7 billion. The smallest models are optimized for on-device inference and can generate audio up to two minutes in length, while the medium and large variants are designed to produce full musical compositions lasting six minutes and 20 seconds—representing a significant extension from Stable Audio 2.0, which was limited to shorter durations.

The company is releasing three of the four models with open weights, enabling researchers and developers to download, use, and modify the code freely. Only the largest model is restricted to API access and paid self-hosting options, with enterprises generating over $1 million in annual revenue required to obtain an enterprise license. This tiered release strategy balances accessibility with commercialization.

Stability AI emphasized that all models in the Stable Audio 3.0 family were trained on fully licensed audio data, following the company's partnership agreements with Warner Music Group and Universal Music Group announced in 2024. This licensing approach distinguishes Stability from competitors facing ongoing litigation over training data sourcing, such as Suno and Udio, which have confronted legal challenges from record labels and publishers over copyright compliance.