thepatch/stable-audio-open-small-GGUF

Stable Audio Open Small GGUF

GGUF conversions for sa3.cpp's optional native stable-audio-tools runtime. This repository is deliberately scoped to Stable Audio Open Small (SAOS). Stable Audio Open 1.0 and Foundation-1 belong in separate model repositories.

The root contains the released ARC-trained SAOS DiT and the shared T5-base encoder and Oobleck decoder. Validated full-checkpoint finetunes are nested under finetunes/:

• finetunes/kickbass/: S3Sound/kickbass, checkpoint kickbass_v1_saos_e257_s18870.ckpt.
• finetunes/jerry-grunge/: thepatch/jerry_grunge, preencoded checkpoint jerry_encoded_bs64_epoch=374-step=3000.ckpt.

Quantization choices

Every component is published at F16, Q8_0, Q5_K_M, and Q4_K_M. Use the same tier for all three components unless you have a reason to mix them.

Tier	Complete bundle	Intended use
F16	1,010 MiB	reference
Q8_0	569 MiB	near-transparent
Q5_K_M	416 MiB	recommended default
Q4_K_M	379 MiB	smallest footprint

Matched 11-second renders validated all four tiers on ARC, KickBass, and Jerry Grunge. Q5_K_M was consistently robust and is only 37 MiB larger than the all-Q4 bundle.

Download and run

SAOS is not part of the default sa3.cpp build. Enable the optional SAT component:

cmake -S . -B build-sat -DSA3_BUILD_SAT=ON -DSA3_CUDA=ON
cmake --build build-sat --config Release --target sat-generate
python tools/download_models.py --sat --sat-model saos --saos-variant arc

Then generate directly from the canonical filenames:

$env:SA3_DEVICE="cuda"
$env:SA3_FLASH_ATTN="1"
sat-generate --model arc --prompt "A short, beautiful piano riff in C minor" `
  --seconds 11 --seed 1234 --out saos.wav

For a finetune, download and select the same variant name:

python tools/download_models.py --sat --saos-variant jerry-grunge
sat-generate --model jerry-grunge `
  --prompt "chillhop synth warm bass dusty drums sidechain 90 bpm" `
  --seconds 11 --seed 4242 --out jerry-grunge.wav

ARC metadata selects ping-pong sampling with 8 steps and CFG 1. The two ordinary rectified-flow finetunes select Euler with 50 steps and CFG 4. --sampler dpmpp --steps 40 --cfg-scale 4 is another validated finetune configuration.

Sources and licensing

• SAOS source: stabilityai/stable-audio-open-small, revision dc620d91535857b72ebb59b4ca45978db6d417f5.
• KickBass source: S3Sound/kickbass, revision 44e51ff50d14f34654a9305580687e9fe875223f. Redistribution here is with direct permission from S3Sound.
• Jerry Grunge source: thepatch/jerry_grunge, revision d23abf859f51af6587da6f1397306d18c8c42edd.
• T5-base source: google-t5/t5-base, revision a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1, under Apache License 2.0.

The Stable Audio weights and derived finetunes remain subject to the Stability AI Community License included as LICENSE.md. T5-base is covered by LICENSE_T5.md. See NOTICE for attribution and modification details. These artifacts rename tensors, serialize them as GGUF, and quantize selected tensors; they do not merge the finetunes or change their model architecture.

Powered by Stability AI.