Stable Audio 3 Core ML

Powered by Stability AI.

Core ML conversion by beshkenadze of Stable Audio 3 Medium and Small-Music for Apple Silicon, macOS 15 and later. No training or fine-tuning was performed. This is a format/runtime conversion, with the original licensing retained.

Source and contents

Source: stabilityai/stable-audio-3-optimized, revision da6edc54ddba10bfd79a077102ded687f80e882b, MLX checkpoints. Original model cards: Medium and Small-Music. Text conditioning derives from Google DeepMind T5Gemma as redistributed by that pinned optimized release; no separate T5Gemma revision is claimed. Stable Audio sampling tables and tokenizer are included.

coreml/v1 contains 74 compiled components: 12 shared text blocks, 24 Medium transformer blocks and 12 decoder blocks, 20 Small-Music transformer blocks and 6 decoder blocks, plus required CPU arrays, tokenizer and sampling tables. Medium uses CPU and Apple Neural Engine together; Small-Music uses CPU. Text conditioning and decoding use CPU. These are not complete standalone pipelines: a compatible sliced Core ML runner is required.

Conversion

Converted the MLX checkpoints to Core ML with computation split into sequential blocks. Medium keeps dense projections eligible for ANE while other arithmetic runs in FP32 on CPU. Removed unused copies of weights and verified that every remaining referenced weight was retained. No audio quality enhancement or new training is claimed. MODIFICATIONS.txt lists the affected binary files. The original conversion release manifest and its per-file checksums are included.

Licenses and attribution

Stable Audio 3: Stability AI Community License (StableAudio-LICENSE.txt). T5Gemma text components: Gemma Terms of Use (Gemma-LICENSE.txt). These files are not Apache- or MIT-licensed weights. By downloading, using or redistributing the applicable files, you agree to these accompanying terms. The Gemma use restrictions in section 3.2, including its Prohibited Use Policy, are enforceable conditions of this distribution, with notice to subsequent users. See NOTICE.txt and the complete license texts. Preserve notices and describe further modifications. No endorsement by Stability AI or Google is implied.

Commercial use/distribution of Stable Audio requires registration with Stability AI; above the applicable aggregate annual revenue limit a separate license is required. See its license for the integrated end-user product exception.

Integrity

manifest.json lists each file's byte size and SHA-256. The v1 paths are immutable; revised conversions must use a new version. Original converted binary files are unchanged from the validated local conversion; added license documents and modification notices do not change their contents.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for beshkenadze/stable-audio-3-coreml

Unable to build the model tree, the base model loops to the model itself. Learn more.