Models¶

demucs-onnx v0.3.0 publishes 9 ONNX repos under the StemSplitio Hugging Face org. All auto-download on first use; nothing else is required.

At a glance¶

Alias	Kind	Stems	fp32 size	fp16weights size	Best for
`htdemucs_ft`	4-specialist bag	drums, bass, other, vocals	1.26 GB (4×316 MB)	660 MB (4×166 MB)	Best SDR. Default.
`htdemucs`	Single file	drums, bass, other, vocals	316 MB	166 MB	Fastest 4-stem startup, ~30% faster than the bag.
`htdemucs_6s`	Single file	drums, bass, other, vocals, guitar, piano	258 MB	136 MB	Need guitar / piano stems. The only ONNX export of the 6-stem variant.
`htdemucs_ft_drums`	Single specialist	drums	316 MB	166 MB	Drum extraction, beat transcription.
`htdemucs_ft_bass`	Single specialist	bass	316 MB	166 MB	Bassline isolation, mix rebalancing.
`htdemucs_ft_other`	Single specialist	other	316 MB	166 MB	Karaoke instrumental (pair with vocals).
`htdemucs_ft_vocals`	Single specialist	vocals	316 MB	166 MB	#1 open-source vocal SDR — vocal removal, acapella, karaoke.

fp16weights files compute in fp32 at runtime — same RAM, same speed, max abs diff vs fp32 weights is ~6 × 10⁻⁵. Use them.

Quality¶

Parity vs PyTorch fp32 (random (1, 2, 343980) input):

Repo	max abs diff	tolerance
`htdemucs-ft-drums-onnx`	1.63 × 10⁻⁴	1 × 10⁻³
`htdemucs-ft-bass-onnx`	1.42 × 10⁻⁴	1 × 10⁻³
`htdemucs-ft-other-onnx`	1.71 × 10⁻⁴	1 × 10⁻³
`htdemucs-ft-vocals-onnx`	1.55 × 10⁻⁴	1 × 10⁻³
`htdemucs-onnx`	6.62 × 10⁻⁴	1 × 10⁻³
`htdemucs-6s-onnx`	2.42 × 10⁻⁴	1 × 10⁻³

Official MUSDB18-HQ SDR (median across 50 test songs, BSS Eval v4):

Model	drums	bass	other	vocals
`htdemucs_ft`	10.11 dB	9.7 dB	7.0 dB	9.19 dB
`htdemucs`	~9.5 dB	~9.0 dB	~5.5 dB	~8.8 dB
`htdemucs_6s`	~9.5 dB	~9.0 dB	~5.5 dB	~8.5 dB

htdemucs_6s has lower vocals/other SDR because the model also predicts guitar and piano — "other" becomes more specific (less of a catchall), trading vocal-recall for guitar/piano clarity.

Full benchmark across every popular open-source separator: StemSplitio/stem-separation-benchmark-2026.

Performance¶

Single 7.8 s segment on Apple M4 Pro (CPU EP):

Mode	Per segment	Per 3-min song	RTF
`separate(model="htdemucs_ft", stems=("vocals",))`	1.6 s	~22 s	0.20
`separate(model="htdemucs_ft")` (bag, 4 stems)	6.4 s	~88 s	0.49
`separate(model="htdemucs")` (single file, 4 stems)	1.6 s	~22 s	0.20
`separate(model="htdemucs_6s")` (single file, 6 stems)	1.6 s	~22 s	0.20

CUDA / DirectML / CoreML EPs are typically ≥ 5× faster on real GPUs.

Programmatic listing¶

from demucs_onnx import list_models
import json
print(json.dumps(list_models(), indent=2))

Or via the CLI:

demucs-onnx list-models

Direct download URLs¶

Every model is downloadable without huggingface_hub. The URLs are stable and CDN-hosted by HF.

# Single-file 4-stem htdemucs (fp16weights variant).
curl -OL https://huggingface.co/StemSplitio/htdemucs-onnx/resolve/main/htdemucs_fp16weights.onnx

# 6-stem variant.
curl -OL https://huggingface.co/StemSplitio/htdemucs-6s-onnx/resolve/main/htdemucs_6s_fp16weights.onnx

# Vocals specialist (the most-used file).
curl -OL https://huggingface.co/StemSplitio/htdemucs-ft-vocals-onnx/resolve/main/htdemucs_ft_vocals_fp16weights.onnx

License¶

All ONNX files are MIT-licensed, matching the original HT-Demucs. Please cite the original authors if you use them in research; see the Skip the infrastructure section on the home page for the bibtex.