---
license: mit
tags:
- audio
- music
- music-source-separation
- demucs
---

# Demucs-mdx

Weights for the `mdx` pretrained model of
[Demucs](https://github.com/adefossez/demucs), music source separation in the
waveform domain. See the [demucs repository](https://github.com/adefossez/demucs)
for how to use them.

This is a bag of 4 model(s), whose outputs are averaged (see `mdx.yaml`
for the per source weights). Each `.safetensors` file contains the weights of one
model, with its class and init arguments as json in the safetensors metadata, and
the full training metadata in the matching `.json` sidecar.

- `0d19c1c6` (`0d19c1c6.safetensors`)
- `7ecf8ec1` (`7ecf8ec1.safetensors`)
- `c511e2ab` (`c511e2ab.safetensors`)
- `7d865c68` (`7d865c68.safetensors`)