pimp-audio-ace-captioner-4bit-v1

PiMP Audio ACE Captioner 4-bit v1

This is a native Swift/MLX model pack derived from ACE-Step/acestep-captioner. It is intended for local, whole-track dataset annotation in PiMP Audio Studio. This is a community conversion and is not an official ACE-Step release.

Pack details

• Task: music captioning and annotation
• Upstream revision: 7109e9af1c6cfe3371730310b3a5cb773145b11e
• Quantization: 4-bit affine, group size 64
• Included model components: thinker and audio tower
• Excluded model components: talker, token-to-wave, and vision
• On-disk size: approximately 4.67 GB
• Runtime: MLX through native Swift code in PiMP Audio Studio

native-model-pack.json is the machine-readable pack manifest. PiMP Audio Studio validates the required file set, byte counts, and SHA-256 checksums before installing the pack.

Availability and use

This repository is packaged for PiMP Audio Studio's Native ACE annotation backend. It is not a standard Transformers checkpoint and is not intended to generate audio.

The model pack is hosted publicly so PiMP Audio Studio can download it without requiring a Hugging Face account. PiMP Audio Studio itself is currently private and has not been released for general download. This repository is therefore not a standalone application or a complete public inference package; the compatible native Swift/MLX runtime currently ships only inside PiMP Audio Studio.

In a compatible PiMP Audio Studio build, open Settings > Models, find Dataset Annotation (Advanced), and click Download beside ACE Captioner 4-bit. The app downloads the pinned revision, verifies every hosted file against the pack manifest, installs it transactionally, and makes it active automatically.

The native annotation backend is unsupported on 16 GB systems, provisional on 24 GB systems, and recommended on systems with at least 32 GB of unified memory.

License and attribution

The upstream model is released under the MIT license. See the ACE-Step captioner model card and the ACE-Step project for the original model, documentation, and attribution.