Acentaur audio-analysis MLX packages
Official public release-distribution repository for the audio-analysis MLX packages used by ACENTAUR 1.0 on Apple Silicon. ACENTAUR Setup downloads only the selected revision-pinned package paths. The packages are derived from the official ACE-Step 1.5 captioner and transcriber checkpoints and retain the upstream MIT license.
These packages are not intended for direct use with LM Studio, Transformers, generic MLX model loaders or Hugging Face Inference Providers. Get ACENTAUR and product information at https://www.einaux.com.
These packages use ACENTAUR's versioned internal audio-analysis contract. Download and installation are managed by ACENTAUR Setup; cloning the complete repository is unnecessary. The original Qwen2.5-Omni wrappers were reduced to the audio tower, dense text model and LM head before conversion. Vision, Talker and Token2Wav weights are not included.
Product packages
| Package | Product role | Precision | Safetensors payload |
|---|---|---|---|
captioner-fast-4bit |
Music description, Fast | 4-bit affine language model, BF16 audio tower | 5,563,709,140 bytes |
captioner-best-bf16 |
Music description, Best | BF16 | 16,510,629,216 bytes |
transcriber-lyrics-8bit |
Lyrics and language transcription | 8-bit affine language model, BF16 audio tower | 9,371,350,953 bytes |
Total Safetensors payload: 31,445,689,309 bytes.
Installer package split
The checkpoints are shipped as two independently selectable Acentaur product packages. Each checkpoint directory is self-contained; neither package loads the original Qwen2.5-Omni weights.
| Acentaur package | Included checkpoint directories | Persistent payload including metadata |
|---|---|---|
Audio Analysis (audio_analysis) |
captioner-fast-4bit, transcriber-lyrics-8bit |
14,967,119,111 bytes |
Detailed Description (audio_analysis_detailed) |
captioner-best-bf16 |
16,526,875,347 bytes additional |
The full union occupies 31,493,994,458 bytes including metadata. Detailed Description depends on Audio Analysis; it adds the full-precision music
captioner and does not replace the compact captioner or lyrics transcriber.
Provenance and conversion
Exact upstream revisions, conversion parameters, file sizes, tensor counts and
SHA-256 values are recorded in manifest.json.
The packages were rebuilt on an M1 Max with the same MLX-VLM 0.6.11 converter code used for the validated M5 reference. Every produced Safetensors shard is byte-identical to the previously tested M5 artifact.
Validation
The deterministic reference matrix reproduced the recorded token hashes:
- captioner 4-bit:
032c00a2a7d42736e1e8aa0b36f3f70fff1b9e139610a26bdedb72286434a95a - transcriber 8-bit:
c85d46f10685dd1c7b450909d0a31b2cbaa9854355b3502c97573e812c5faded
The browser Analysis Lab remains the human regression surface. A normal Acentaur installation only needs the three packages above and the bundled MLX runtime; it never downloads the original 22.36-GB Omni checkpoints.
Public, anonymous revision-pinned download is part of the ACENTAUR 1.0 Setup and release contract. Changing the repository to private would be a breaking distribution change and requires a new authenticated delivery contract.