Stable Audio 3¶

Stable Audio 3 small generates stereo 44.1 kHz audio from a text prompt: a T5Gemma text encoder, a 20-layer diffusion transformer sampled with ping-pong in 8 steps, and the SAME-S decoder. It generates up to 120 seconds.

Model	Purpose
`mlx-community/stable-audio-3-small-sfx`	Sound effects
`mlx-community/stable-audio-3-small-music`	Music

Generate¶

CLIPython

python -m mlx_audio.music.generate \
    --model mlx-community/stable-audio-3-small-sfx \
    --caption "Rain on a tin roof with distant thunder" \
    --duration 10 \
    --seed 7 \
    --output rain.wav

from mlx_audio.music import load

model = load("mlx-community/stable-audio-3-small-music")

for result in model.generate(
    text="Lofi house loop, 120 BPM",
    duration=20,
    steps=8,
    seed=7,
):
    print(result.audio.shape, result.sample_rate)

--cfg-scale defaults to 1, which turns guidance off. A larger scale doubles the model calls per step and enables --negative-prompt.

License¶

The code follows Stability AI's MIT-licensed optimized/mlx code, Copyright (c) 2026 Stability AI. The weights are under the Stability AI Community License and the text encoder under the Gemma Terms of Use. Read both before you use or redistribute the model.