mlx-musicgen-style

mlx-musicgen-style (MusicGen Style, MLX)

Non-commercial licence. These weights are CC-BY-NC-4.0 and may not be used for commercial purposes. See Licence below.

MLX conversion of facebook/musicgen-style for inference on Apple silicon via mlx-audiogen.

Licence

Upstream MusicGen is released under two different licences, and it matters which applies to what:

• Model weights — CC-BY-NC-4.0 (non-commercial). That is what this repository contains, so that is what governs it.
• The upstream code is MIT. That does not extend to these weights.

Key terms, summarised (the licence text governs):

• Free to share and adapt for non-commercial purposes.
• You must give appropriate credit and indicate if changes were made.
• No commercial use — including use primarily directed towards commercial advantage or monetary compensation.

mlx-audiogen's own Apache-2.0 licence covers the conversion code only. The full licence is in LICENSE; attribution and the statement of modifications are in NOTICE.

Attribution

MusicGen Style was created by Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi and Alexandre Défossez, and published in Simple and Controllable Music Generation.

Modifications

These weights are a modified version of the upstream checkpoint, converted to the MLX array format for inference on Apple silicon. Specifically: convolution weights are transposed where MLX's layout differs from PyTorch's, tensors are re-serialised as safetensors, weight keys are remapped to match this project's module names, and the text encoder is split into a separate file. The model is re-implemented against MLX rather than PyTorch.

No weights are retrained or fine-tuned.

Tensor precision as shipped: the decoder and the style conditioner store float16; the MERT feature extractor and the T5 text encoder store float32. Upstream distributes an audiocraft state_dict.bin rather than safetensors, so the float16 tensors may be a precision reduction relative to it — treat them as possibly rounded. No weights are retrained or fine-tuned either way.

Bundled components

• T5 text encoder — https://huggingface.co/google-t5/t5-base, Apache-2.0, Google LLC. Licence text: LICENSE.t5-apache-2.0
• MERT-v1-95M audio feature extractor — https://huggingface.co/m-a-p/MERT-v1-95M, CC-BY-NC-4.0, Music Audio Pretrain (m-a-p). Licence text: LICENSE

Where component licences differ, the most restrictive term governs this repository: CC-BY-NC-4.0.

Usage

pip install "mlx-audiogen[server]"
mlx-audiogen --model musicgen --prompt "happy rock song" --seconds 5 \
  --weights-dir musicgen-style

The weights download automatically on first use.