Stability AI has released Stable Audio 3.0, a family of AI music generation models that brings full song composition to smartphones and laptops for the first time. The release includes three open-weight models that users can download and customize, plus a larger enterprise-focused model available through the company’s API.
The move represents a significant shift in AI music generation, making professional-quality audio creation accessible on consumer devices while addressing the licensing concerns that have plagued the industry. Unlike many competing models that restrict commercial use or carry legal risks, Stable Audio 3.0 is trained exclusively on licensed data and grants users full ownership of their generated content.
Four models for different use cases
The Stable Audio 3.0 family consists of four models designed for specific deployment scenarios:
- Stable Audio 3.0 Small SFX: Generates sound effects on mobile devices and consumer laptops
- Stable Audio 3.0 Small: Creates complete music compositions on-device
- Stable Audio 3.0 Medium: Offers improved musical structure and can generate tracks up to 6 minutes and 20 seconds
- Stable Audio 3.0 Large: The most advanced model, designed for music platforms requiring high-volume, low-latency generation
The breakthrough here is enabling full music composition on portable devices. Previous versions of Stable Audio Open could only generate 11-47 second clips, while the new Small model creates complete two-minute tracks entirely offline on consumer hardware.
Commercial licensing and ownership rights
Stability AI addresses one of the biggest pain points in AI-generated music: legal uncertainty. All Stable Audio 3.0 models use fully licensed training data, and users retain complete ownership of their outputs under the Stability AI Community License.
Organizations with annual revenue exceeding $1 million can access commercial coverage through an Enterprise license, which includes legal indemnification. This approach contrasts sharply with other open music models that either restrict commercial use or carry risks from unlicensed training data.
Technical advances enable longer, more flexible generation
The models introduce several technical improvements that expand creative possibilities:
- Variable-length generation: Create audio at per-second granularity, up to six minutes for Medium and Large models
- On-device composition: Generate complete musical works without internet connectivity
- LoRa training support: Customize models using your own music library with efficient fine-tuning methods
- Audio inpainting: Edit specific segments, rework parts of songs, or extend compositions without starting over
These capabilities stem from a new semantic-acoustic autoencoder architecture that Stability AI has documented in a research paper accompanying the release.
Industry partnerships and future plans
The release comes as Stability AI works to position itself as the responsible choice in AI music generation. The company has established partnerships with major labels including Universal Music Group and Warner Music Group, signaling industry acceptance of its licensing approach.
Stability AI is also developing a suite of products specifically for musicians, with early access available through a waitlist. This suggests the company sees professional music creation, not just hobbyist experimentation, as a key market.
The three smaller models are immediately available for download on Hugging Face, while the Large model requires API access or enterprise deployment. The models will also be supported on platforms like ComfyUI, expanding their accessibility to the broader AI creative community.
