---
license: mit
tags:
  - music-source-separation
  - stem-separation
  - demucs
  - ggml
  - webassembly
---

# htdemucs 4-source weights, ggml format

HT-Demucs (Demucs v4) weights converted to the ggml binary format used by
[demucs.cpp](https://github.com/sevagh/demucs.cpp), for running music source
separation in WebAssembly.

Re-hosted so [Audio Magic](https://audiomagic.studio) has a pinned copy it
controls. **Nothing has been modified** — byte-identical to the upstream
conversion.

## Provenance

| | |
|---|---|
| Original model | [Demucs v4 / HT-Demucs](https://github.com/adefossez/demucs) — Meta Platforms, MIT |
| Inference engine | [demucs.cpp](https://github.com/sevagh/demucs.cpp) — Sevag Hanssian, MIT |
| ggml conversion | [`Retrobear/demucs.cpp`](https://huggingface.co/datasets/Retrobear/demucs.cpp), revision `8f58ac0491bbea657275bcd3e38af1ab3a27bfc9` |
| File | `ggml-model-htdemucs-4s-f16.bin` → `ggml-htdemucs-4s-f16.bin` |
| Size | 83,994,361 bytes |
| SHA-256 | `72b17c42d308982ddb5069bc3bf48b81a5aac4cb6516e4366c0fa7cef6df0064` |

Stems, in output order: **drums, bass, other, vocals**. Weights are f16.

## Verification

Loaded into a `demucs.cpp` WebAssembly build and run against a real
44.1 kHz stereo recording: the model initialises in 0.2 s and the four stems
reconstruct the input at **28.5 dB SNR**.

## Licence

Demucs and its weights are MIT (Meta Platforms); demucs.cpp is MIT
(Sevag Hanssian). Both notices are included as `LICENSE-demucs` and
`LICENSE-demucs.cpp`.

## Citation

```bibtex
@inproceedings{rouard2023hybrid,
  title     = {Hybrid Transformers for Music Source Separation},
  author    = {Rouard, Simon and Massa, Francisco and D{\'e}fossez, Alexandre},
  booktitle = {ICASSP 23},
  year      = {2023}
}
```