Demucs for MLX
MLX weights for all eight Demucs music source separation models, for use on Apple silicon with:
demucs-mlx— Python package and CLIdemucs-mlx-swift— Swift package and CLI for macOS and iOS
Provenance
These files were converted by ssmall256 from Meta's official Demucs checkpoints using the demucs-mlx converter (python -m demucs_mlx.mlx_convert <model>). The converter loads each official checkpoint with PyTorch's restricted weights_only loader, checks it against the Demucs registry's checksums, and writes safetensors — no pickle files are produced or needed.
Every member of every model was compared against its PyTorch source on the same input before release; agreement is between 75 and 120 dB signal-to-noise. The tensors are float32. Converting again with demucs-mlx reproduces these files byte for byte, so you can check them yourself against the digests below.
Models
| Model | Description | Stems | Members | Size | SHA-256 of .safetensors |
|---|---|---|---|---|---|
htdemucs |
Hybrid Transformer Demucs (default) | 4 | 1 | 160 MB | 339d267a7a6983a11eedbdc00413c602a65e9b9103f695fb5c2b2a481cd9d297 |
htdemucs_ft |
Fine-tuned HTDemucs, one specialist per stem | 4 | 4 | 641 MB | 53f03b1ad4b4d211025a35da65460ba61a17547adf9c0544cad0ebcc8d7bbabb |
htdemucs_6s |
HTDemucs with piano and guitar | 6 | 1 | 105 MB | d298f7f746bf53c21baad44fb08e88807ef47feb551dd22f1601a546c85b8e02 |
hdemucs_mmi |
Hybrid Demucs v3 | 4 | 1 | 319 MB | 39f359110433930c2a589131f84c03c26bbd209e89e10e6abbcb6062c131debc |
mdx |
MDX challenge track A bag | 4 | 4 | 1,318 MB | c95dab261c766fc50caadcd047aa2c125759b9b48efdac3f2aaab0fa35d8c41f |
mdx_q |
mdx from the DiffQ-quantized checkpoints |
4 | 4 | 1,318 MB | d7f31edb6b37b5ee391d104e1f88cb70c56ca82e3f6f9e8f4f3cd2df6e9bddfc |
mdx_extra |
MDX challenge track B bag | 4 | 4 | 1,276 MB | d1c969aa0a69417e767b23f97d10df944a4c9883febe853ab6d91ad0cb2276fe |
mdx_extra_q |
mdx_extra from the DiffQ-quantized checkpoints |
4 | 4 | 1,276 MB | 82310bf4d1f32b8044cba6c192af77ab8d12a0acdedd7bf841caa78a61bd5839 |
Four-stem models produce drums, bass, other and vocals; htdemucs_6s adds guitar and piano. The _q names identify the original quantized checkpoints; the weights here are ordinary float32.
Each model is two files: <model>.safetensors and <model>_config.json. The config (format version 1) records the model classes and constructor arguments, ensemble weights, the official checkpoint signatures, and the SHA-256 of the weights. Both loaders verify that digest before building a model. This two-file layout is the shared cache format: demucs-mlx keeps it in ~/.cache/demucs-mlx (or DEMUCS_MLX_CACHE_DIR), and other tools can read the same directory instead of storing a second copy.
Use
To download one model, or everything, into a directory yourself:
hf download ssmall256/demucs-mlx htdemucs.safetensors htdemucs_config.json --local-dir ~/.cache/demucs-mlx
hf download ssmall256/demucs-mlx --local-dir ~/.cache/demucs-mlx
Python — demucs-mlx downloads the model it needs on first use and checks it against these digests, so there is nothing to fetch by hand:
pip install demucs-mlx
demucs-mlx song.wav
Swift — in demucs-mlx-swift, Separator.load() downloads and verifies the model on first use, and shares the same cache directory:
let separator = try await Separator.load()
let result = try await DemucsAudio.separate(songURL, using: separator)
try DemucsAudio.export(result, to: outputDirectory)
License and credit
MIT, the same license as Demucs. The models were trained by Alexandre Défossez and collaborators at Meta AI Research:
- Défossez, Hybrid Spectrogram and Waveform Source Separation, 2021.
- Rouard, Massa, Défossez, Hybrid Transformers for Music Source Separation, ICASSP 2023.
The MLX conversion, the demucs-mlx converter and both inference implementations are by ssmall256.
Quantized