mlx_conversion.json

{
  "source": "OpenMOSS-Team/MOSS-Transcribe-Diarize",
  "format": "mlx-safetensors",
  "model_type": "moss_transcribe_diarize",
  "quantization": {
    "bits": 8,
    "group_size": 64,
    "mode": "affine",
    "scope": "text_backbone_linears_only",
    "excluded_prefixes": ["model.whisper_encoder", "model.vq_adaptor"],
    "excluded_layers": ["model.language_model.embed_tokens"],
    "note": "Token embedding kept full precision: mlx-audio-swift loads it as a plain Embedding and does not swap quantized embeddings, so a quantized embed_tokens fails its load-time shape check."
  }
}