mlx-community/demucs-mlx-fp16

Originally from: iky1e/demucs-mlx-fp16

Float32 variant: mlx-community/demucs-mlx

Demucs — MLX (float16)

Float16 MLX-compatible weights for all 8 pretrained Demucs models, converted to safetensors format for inference on Apple Silicon.

This is the float16 variant of iky1e/demucs-mlx — same models, half the file size, identical output quality. Recommended for Apple Silicon where memory is constrained (iOS, smaller Macs).

Demucs is a music source separation model that splits audio into stems: drums, bass, other, vocals (and guitar, piano for 6-source models).

Models

Model	What it is	Architecture	Sub-models	Sources	Weights (fp16)	Weights (fp32)
htdemucs	Default v4 model, best speed/quality balance	HTDemucs (v4)	1	4	80 MB	160 MB
htdemucs_ft	Fine-tuned v4, best overall quality	HTDemucs (v4)	4 (fine-tuned)	4	321 MB	641 MB
htdemucs_6s	6-source v4 (adds guitar + piano stems)	HTDemucs (v4)	1	6	52 MB	105 MB
hdemucs_mmi	v3 hybrid, trained on more data	HDemucs (v3)	1	4	160 MB	319 MB
mdx	v3 bag-of-models ensemble	Demucs + HDemucs	4 (bag)	4	659 MB	1.3 GB
mdx_extra	v3 ensemble trained on extra data	HDemucs	4 (bag)	4	638 MB	1.2 GB
mdx_q	Quantized v3 ensemble (same quality, smaller)	Demucs + HDemucs	4 (bag)	4	659 MB	1.3 GB
mdx_extra_q	Quantized v3 extra ensemble	HDemucs	4 (bag)	4	638 MB	1.2 GB

All models output stereo audio at 44.1 kHz.

Float16 vs Float32

Output quality is identical — max sample difference is 3.1e-5 (one int16 LSB), correlation > 0.999999999. MLX on Apple Silicon upcasts float16 weights to float32 for computation, so the math is the same.

Metric	float32 (iky1e/demucs-mlx)	float16 (this repo)
htdemucs file size	160 MB	80 MB
htdemucs RSS (peak memory)	1311 MB	1210 MB
htdemucs speed (M1 Pro)	7.1s	7.9s
Output quality	reference	identical

Origin

• Original model/repo: adefossez/demucs
• Float32 weights: iky1e/demucs-mlx
• License: MIT (same as original Demucs)
• Conversion path: PyTorch checkpoints → safetensors float32 → float16
• Swift MLX port: kylehowells/demucs-mlx-swift

Files

Each model consists of two files at the repo root:

• {model_name}.safetensors — model weights (float16)
• {model_name}_config.json — model class, architecture config, and bag-of-models metadata

Usage

Swift (demucs-mlx-swift)

Point the model directory or repo to this float16 variant:

# Use float16 models from local directory
demucs-mlx-swift -n htdemucs --model-dir /path/to/demucs-mlx-fp16 song.wav

# Or set the HF repo environment variable
export DEMUCS_MLX_SWIFT_MODEL_REPO=iky1e/demucs-mlx-fp16
demucs-mlx-swift -n htdemucs song.wav

Or use the Swift API directly:

import DemucsMLX

let separator = try DemucsSeparator(modelName: "htdemucs")
let result = try separator.separate(fileAt: URL(fileURLWithPath: "song.wav"))

Converting from PyTorch

To reproduce the export directly from PyTorch Demucs checkpoints:

pip install demucs safetensors numpy

# Export all 8 models as float16 (default)
python export_from_pytorch.py --out-dir ./output

# Export as float32
python export_from_pytorch.py --out-dir ./output --dtype float32

The conversion script (export_from_pytorch.py) is available in the demucs-mlx-swift repo under scripts/.

Citation

@inproceedings{rouard2022hybrid,
  title={Hybrid Transformers for Music Source Separation},
  author={Rouard, Simon and Massa, Francisco and Defossez, Alexandre},
  booktitle={ICASSP 23},
  year={2023}
}

@inproceedings{defossez2021hybrid,
  title={Hybrid Spectrogram and Waveform Source Separation},
  author={Defossez, Alexandre},
  booktitle={Proceedings of the ISMIR 2021 Workshop on Music Source Separation},
  year={2021}
}