demucs-htdemucs-mlx

demucs-htdemucs-mlx (HTDemucs v4, MLX)

MLX conversion of facebookresearch/demucs for inference on Apple silicon via mlx-audiogen.

Licence

Meta releases Demucs under the MIT licence. Its README states it plainly — "Demucs is released under the MIT license as found in the LICENSE file" — and that is the whole of what upstream documents.

MIT is permissive, but it is not obligation-free: the licence requires that

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

which is why this repository ships the full LICENSE rather than only a license: mit tag.

Before commercial use, do your own diligence. Upstream states one licence for the project and publishes no separate terms for the released checkpoints — the licence text itself grants rights in "the Software". Separately, the README records that htdemucs was trained on MUSDB18-HQ plus 800 extra songs, and MUSDB18-HQ carries its own non-commercial terms. Whether training-data terms reach model weights is unsettled, and Meta evidently takes the position that MIT covers what it ships. This card reports what upstream states rather than resolving that question for you.

mlx-audiogen's own Apache-2.0 licence covers that project's code — it grants you no rights in these weights. The full licence is in LICENSE; attribution and the statement of modifications are in NOTICE.

Attribution

HTDemucs v4 was created by Simon Rouard, Francisco Massa and Alexandre Défossez, and published in Hybrid Transformers for Music Source Separation.

Modifications

These weights are a modified version of Meta's HTDemucs v4 checkpoint, converted to the MLX array format for inference on Apple silicon. Specifically: the upstream pickled Torch .th checkpoint is read for its tensor state dict only, weight keys are remapped to match this project's module names, convolution weights are transposed where MLX's layout differs from PyTorch's (both the 1-D and 2-D cases), fused QKV projections are split into separate Q, K and V tensors, and the result is re-serialised as safetensors. The model is re-implemented against MLX rather than PyTorch.

No weights are retrained or fine-tuned.

Tensor precision is unchanged: the upstream checkpoint stores float32 and so does this conversion. No weight value is altered by rounding — the Demucs converter performs no dtype cast.

Usage

from mlx_audiogen.models.demucs.pipeline import DemucsPipeline

# Auto-downloads weights from this repo
pipeline = DemucsPipeline.from_pretrained()

stems = pipeline.separate(audio_np, sample_rate=44100)
# stems = {"drums": np.ndarray, "bass": ..., "other": ..., "vocals": ...}

The weights download automatically on first use.

Provenance

The checkpoint this was converted from is distributed by Meta at https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/955717e8-8726e21a.th, referenced by the upstream project above.