Stable Audio 3.0

What it is

Stable Audio 3.0 is a family of audio generation models released by Stability AI around 20 May 2026, together with an arXiv paper titled "Stable Audio 3" dated 18 May 2026 ^[1][2]. It is not a single network but four models of different sizes, aimed at both music and sound effects ^[1]. The paper describes Stable Audio 3 as a family of fast latent diffusion models built for variable-length audio generation and editing ^[4].

The release also includes editing rather than generation alone: the models can work on audio that already exists, not only create new clips from a text prompt ^[1].

How it works

The models do not run diffusion directly on the raw waveform. According to the paper, they operate on a semantic-acoustic autoencoder that projects audio into a compact latent space, preserving fidelity while encouraging the representation to carry semantic structure ^[4]. Latent diffusion means noise is added and removed inside that compressed representation instead of in the sample-by-sample audio signal, which makes generation far cheaper.

Two design choices are called out in the paper. Adversarial post-training was used to cut the number of inference steps required, while also improving fidelity and how closely the output follows the prompt ^[4]. That shows up as speed: the models generate music and sounds in less than two seconds on an H200 GPU, and in a few seconds on a MacBook Pro M4 laptop ^[4].

The four models

Small SFX and Small each have 459 million parameters; Medium has 1.4 billion and Large 2.7 billion ^[1]. Small SFX, Small and Medium are released as open-weight models, while Large is offered through the Stability AI API and through enterprise self-hosting ^[1]. The paper releases the weights of the small and medium models together with their training and inference pipeline ^[4].

The size split lines up with a split in use. The two Small models generate up to 2 minutes of audio and are intended for on-device use on phones and laptops; Medium and Large reach up to 6 minutes 20 seconds and are built to keep musical structure over that length ^[1].

What it can do

Beyond generating music and sound effects from a prompt, the family supports audio inpainting, continuation and variable-length generation with per-second control ^[1]. Inpainting means filling a gap inside existing audio; continuation means extending a clip forwards; per-second control means the generated length can be set second by second rather than only in fixed blocks.

Licensing and training data

Stability says all Stable Audio 3.0 models were trained on fully licensed data ^[1]. The models are released under the Stability AI Community License, under which organisations with more than $1 million in annual revenue need an enterprise license ^[1][3].

Where it sits

The way the family is divided keeps two kinds of use apart. The smaller models are published as open weights and are meant to run locally on a device, while the largest sits behind an API and enterprise self-hosting, with the revenue threshold in the community licence marking the boundary between free and commercial use at scale ^[1][3]. Because the paper sets out the training and inference pipeline alongside the released weights, the open part of the family comes with a documented route from training to deployment ^[4].