Demucs (HTDemucs)

Meta AI Research

The separation model everything else gets benchmarked against.

In short

Demucs, in its hybrid transformer form HTDemucs, is Meta Research's source separation model and the highest-quality open option for splitting a mix into vocals, drums, bass and other. It is not a generation model.

Demucs (HTDemucs) specifications

Output: 4 stems: drums, bass, vocals, other
Quality: ~9.0–9.2 dB SDR (MUSDB HQ)
Architecture: Hybrid transformer, time domain
Weights: Open
Runtime: GPU recommended
Status: Evaluating

Figures verified 2026-08-10. This field moves quickly — re-check before relying on them.

About Demucs (HTDemucs)

Separation is a different problem from generation and Demucs is the reference solution to it. Given a finished stereo mix it recovers the individual sources, and it does so with noticeably fewer artefacts than the alternatives — particularly on transients, where lesser models smear drum hits into a wash. That matters because transient damage is what makes separated stems unusable in a remix: you can work around a slightly dull vocal, but you cannot work around drums that have lost their attack. Almost every commercial stem-splitting service is running Demucs or something derived from it, which is worth knowing before paying a subscription for what is freely available. Running it locally is straightforward, needs no exotic hardware, and keeps unreleased material private. Active maintenance means the quality has continued improving rather than freezing at publication.

Strengths

What Demucs (HTDemucs) is good at

Best public separation quality

Fewer artefacts than any other open option, which is the whole game in separation — a stem you can hear the algorithm working on is a stem you cannot use.

Preserves transients well

Drum attacks survive separation intact, and transient damage is the specific failure that makes separated stems useless in a remix.

Actively maintained

Continued maintenance means quality has improved since publication rather than freezing, which is unusual for a research release.

Demucs (HTDemucs) limitations

Large model, needs real compute
Cannot run in a browser tab
Four fixed stems

How it compares

Choosing Demucs (HTDemucs)

What comes back

Demucs, in its hybrid transformer form HTDemucs, is Meta Research's source separation model and the highest-quality open option for splitting a mix into vocals, drums, bass and other. It is not a generation model.

Compare with MusicGenerate

Demucs (HTDemucs) review, pricing and alternatives

Searched as demucs, htdemucs, demucs stem separation, demucs github and best stem separation. The last query is the honest one — this is the reference the paid services are built on.

Demucs (HTDemucs) — questions people ask

Is Demucs the best stem separation available?

It is the highest-quality open option and the reference the others are benchmarked against. Most commercial stem-splitting services are running Demucs or something derived from it.

Why do Demucs stems sound better than other tools?

Fewer artefacts, and particularly better transient preservation. Transient damage is what makes separated drums unusable in a remix — you can work around a dull vocal, not around drums that lost their attack.

Can I run Demucs on a normal computer?

Yes. It runs locally without exotic hardware, which also keeps unreleased material off someone else's server.

Does separating a track give me the right to release it?

No. Separation creates no new rights. For private practice and study it is fine; a public release needs clearing with the rights holder of the original recording.

Is Demucs free?

Yes, it is open source and free to run locally. Most commercial stem-splitting services charge for what is essentially a hosted wrapper around it.

How long does Demucs take to separate a song?

A few minutes on CPU and considerably less on a GPU, depending on the model variant and the shift settings you choose — those trade processing time for quality.

Keep exploring