Stability AI drops open-weight Stable Audio 3.0

Stable Audio 3.0 is a new family of music and sound generation models you can actually download and run locally. Put simply, it lets you generate complete, structured songs of up to 6 minutes and 20 seconds, or spit out instant sound effects directly on your personal computer and mobile devices.

The real highlight here is the combination of open weights and clean data. Stability AI released four models and made the weights for three of them open on Hugging Face. Crucially, the entire lineup was trained on fully licensed audio, meaning you own your outputs and can legally commercialize them under the Stability AI Community License if your annual revenue is under $1 million.

The lineup and what it can do

The 3.0 family breaks down into four clear tiers tailored to different compute setups.

At the lighter end sit Stable Audio 3.0 Small SFX and Small, both weighing in at 459 million parameters. They are optimized for on-device inference on standard laptops and smartphones without an internet connection. Small SFX handles Foley and sound design, while Small is capable of composing complete musical tracks up to two minutes long. For reference, the previous Stable Audio Open was capped at 47 seconds, and its tiny variant only managed 11 seconds.

If you want richer instrumentation and longer structures, Stable Audio 3.0 Medium comes with 1.4 billion parameters. It cranks generation length up to 6 minutes and 20 seconds with noticeable gains in melodic coherence and musical phrasing.

The flagship 2.7 billion parameter Large model delivers the highest audio fidelity and lowest latency, but Stability AI is keeping its weights closed. It is strictly available via the company API or paid enterprise self-hosting.

Under the hood and how to run it

Generating coherent music past the six-minute mark without structural collapse relies on Stability's new semantic-acoustic autoencoder architecture.

Instead of forcing rigid audio chunks, Stable Audio 3.0 supports variable-length generation with per-second granularity. If you need a 43-second loop for a game level, you get exactly 43 seconds.

It also brings image-style inpainting to audio workflows. You can perform single-segment or multi-segment editing to replace an awkward bridge or solo, as well as causal continuation to naturally extend a track past its original ending.

For builders and sound designers, Stability AI is publishing official LoRA fine-tuning documentation alongside the weights for the Small and Medium tiers. You can adapt the models to your own private sample libraries using relatively modest VRAM.

You can grab the open weights for Small SFX, Small, and Medium right now on Hugging Face, with integrations rolling out across tools like ComfyUI. Those who prefer cloud generation can tap the official API.

Why it matters

For indie game devs, video creators, and musicians, Stable Audio 3.0 solves the biggest headache in generative audio: copyright anxiety. While competitors like Suno and Udio fight major labels in court, Stability AI trained this suite on licensed material through deals with Universal Music Group and Warner Music Group, giving anyone under the $1M revenue bar a clear runway to monetize.

Running full-length song composition directly on consumer hardware is a genuine milestone for offline creative tools. While raw audio fidelity on the 459M models naturally won't match a dedicated cloud cluster, the ability to fine-tune with LoRA, inpaint tracks locally, and avoid subscription lock-in makes this a must-try foundational toolkit for the open audio ecosystem.