Music Tech8 min read

Demucs vs Spleeter: 9.0 dB to 5.9, and the Model to Pick

Demucs vs Spleeter, re-read 16 Nov 2026: Demucs v4 scores 9.0 dB SDR to Spleeter's 5.9. The best Demucs model, why there is no v5, and a free alternative.

Published Updated By Veena Studio

Demucs vs Spleeter is not a close race on quality: in its authors' own table, Hybrid Transformer Demucs scores 9.0 dB SDR on the MUSDB HQ test set to Spleeter's 5.9. Spleeter's one win is speed, "100x faster than real-time" on a GPU. Veena, a free AI DAW with a personal music producer built in, skips the choice: it splits any track you upload into four stems in your browser, with nothing to install.

Most separation tools are wrappers around a handful of models. Knowing which is which tells you what quality to expect. Every figure below comes from the Demucs and Spleeter READMEs on GitHub and their PyPI pages, re-read on 16 November 2026.

Spleeter vs Demucs: the short answer

  • Quality: Demucs. 9.0 dB SDR for the fine-tuned v4 model against 5.9 for Spleeter, in the table on the Demucs README.
  • Speed: Spleeter. Deezer's README promises 4 stems "100x faster than real-time when run on a GPU". The Demucs README puts its CPU time at "roughly equal to 1.5 times the duration of the track".
  • Upkeep: Demucs. Its author's fork shipped Demucs 4.1.0 on 11 July 2026 and runs on current Python (3.10 or later). Spleeter's latest release on PyPI is 2.4.2, from 3 April 2025, and it installs only on Python below 3.12.
  • No setup at all: Veena. Upload a song, right-click it, and four stems land on your timeline in about 45 to 120 seconds.

Demucs vs Spleeter at a glance

November 2026VeenaDemucs v4 (run it yourself)Spleeter (run it yourself)
What you installNothing: a browser tabPython 3.10 or later; runs on CPU, far faster with a GPUPython 3.8 to 3.11, TensorFlow and ffmpeg; "known issues with Apple M1 chips"
How you splitRight-click a clip and choose "Separate into Stems", or ask CoProducerA command line, such as demucs -n htdemucs_ft song.mp3A command line, such as spleeter separate -p spleeter:4stems
Where the stems landOn new tracks in your project, lined up with the originalAudio files in a folderAudio files in a folder
What you do nextKeep producing: CoProducer builds around a stem or turns it into MIDIOpen another appOpen another app

Which Demucs model is best?

The short answer: htdemucs_ft. The Demucs README lists its pretrained models, which you pick with the -n flag:

  • htdemucs — the default: the first Hybrid Transformer model, "Trained on MusDB + 800 songs".
  • htdemucs_ft — the fine-tuned version: "separation will take 4 times more time but might be a bit better". The pick when quality matters more than time.
  • htdemucs_6s — six sources, adding guitar and piano as their own stems; the README itself warns that "the piano source is not working great at the moment".
  • hdemucs_mmi — Hybrid Demucs v3, retrained on MusDB plus 800 songs.
  • mdx and mdx_extra — the 2021 Music Demixing Challenge models; mdx_q and mdx_extra_q are quantized versions with a smaller download, though "quality can be slightly worse".

The short version: htdemucs_ft for a final split, htdemucs when you need speed, and htdemucs_6s only when you need a guitar stem.

Is there a Demucs v5? No. The newest release on 16 November 2026 is Demucs 4.1.0 (PyPI, 11 July 2026), which its README describes as "modernized packaging, lighter inference dependencies, numerous fixes, and pretrained models hosted on Hugging Face". The models are still the v4 family, so "Demucs v4" and "htdemucs v4" are the same thing. The paper's stronger sparse-attention model, at 9.20 dB, is not in the release: it "requires custom CUDA code that is not ready for release yet".

Spleeter (2019)

Deezer's release that made separation mainstream: a library "with pretrained models written in Python" that "uses Tensorflow". A spectrogram-based U-Net predicts a mask for each source, with models for 2 stems (vocals and accompaniment), 4 stems (vocals, drums, bass, other) and 5 stems (adding piano).

Strengths: fast, light, easy to run, still fine for rough work — Deezer says it separates "to 4 stems 100x faster than real-time when run on a GPU". Weaknesses: noticeable artifacts, especially watery vocals. It works on magnitude spectrograms and reuses the original phase, which is a significant part of the artifact, and it scores 5.9 dB SDR in the Demucs authors' comparison.

Still widely embedded in older tools.

Demucs (2019 onward)

The model family its author built at Meta, and the significant advance: it started out working directly on the waveform rather than the spectrogram.

That matters because waveform models handle phase natively rather than reconstructing it. The result sounds meaningfully more natural — less of the phasey quality that characterises spectrogram separation.

Hybrid Demucs (v3) added a spectrogram branch alongside the waveform branch, taking advantage of both. Hybrid Transformer Demucs (HTDemucs, v4), released in November 2022, added Transformer layers for better long-range context; its README says it "achieves a SDR of 9.00 dB on the MUSDB HQ test set".

Strengths: the highest-scoring widely available open model in its authors' own table, and natural-sounding output. Weaknesses: computationally heavier and slow on CPU without a GPU. The original Meta repository now says it "is not maintained anymore", and the copy its README calls "the officially maintained Demucs" is its author's fork at github.com/adefossez/demucs, which warns: "expect slow replies and no new feature for now".

Band-split approaches

Newer models split the spectrum into bands and process each with dedicated subnetworks, recognising that vocals, bass and cymbals need different treatment. Strong results, particularly on vocals: in the Demucs authors' own table, Band-Split RNN reaches 8.2 dB SDR without extra data and 9.0 with it.

The research has moved to the challenges: the Sound Demixing Challenge 2023, run on AIcrowd with Sony, Moises.AI and Mitsubishi Electric Research Laboratories, had a music track that splits a song into vocals, bass, drums and other, and its winners' models are listed in the challenge papers (AIcrowd, read 16 November 2026).

Open-Unmix

An open reference implementation designed for research reproducibility, scoring 5.3 dB SDR in the same table. Not the best-sounding, but valuable as a well-documented baseline.

What commercial tools use

Most separation products build on Demucs variants or proprietary models trained on similar principles, often with:

  • Larger private training sets — the single biggest quality factor
  • More stems — separating guitar and piano rather than lumping them into "other"
  • Post-processing to reduce artifacts
  • Server GPUs for speed

The differences between good commercial tools are usually smaller than their marketing suggests, because they share architectural lineage.

Running it yourself

Demucs is open source, free, and runs locally — which also means your audio never uploads. The maintained fork needs Python 3.10 or later, and its README's quickest route is a single command, uvx demucs MY_TRACK.mp3. It works far faster with a GPU, but CPU works.

For anyone processing a lot of material, or working with confidential audio, this is the option worth knowing about. Just know you are adopting research code whose author has said there will be no new features.

The Demucs alternative with nothing to install

Veena, the AI DAW with a personal music producer built in, has a free stem splitter on every plan. Right-click any audio clip and choose "Separate into Stems", or ask CoProducer. A full song takes about 45 to 120 seconds, and the four stems — vocals, drums, bass and other — land on new tracks lined up with the original, which is muted so nothing doubles. Separating by right-click doesn't use any of your daily CoProducer allowance.

Need just one part? Isolate mode gives you one stem plus the rest of the mix, such as a vocal and an instrumental. Files up to 95 MiB work, about eight and a half to nine and a half minutes of stereo audio. The step-by-step lives in the guide to stem separation in Veena.

The practical takeaway

Separation quality across good tools is closer than it looks. The differentiator is usually what happens next: whether you get four files in a folder, or stems in a project you can immediately edit. In Veena, the next step is the producer's: ask CoProducer to build a new beat under your vocal, or turn the bass stem into MIDI you can rewrite note by note.

Related reading: stem splitter, best stem separation tools, AI music tools and stem bleed.

Try this in Veena

Split my uploaded demo into stems, keep my vocal, and build a new 120 BPM house groove in A minor under it: four-on-the-floor drums with open hats, a rolling bassline, warm chord stabs and a soft pad. 8-bar intro, 16-bar verse, 16-bar chorus.

Open Veena free

Opens a new project in Veena, and CoProducer starts building from this prompt.

Frequently asked questions

Spleeter vs Demucs: which is better?

Demucs, for quality. Its authors' table puts the fine-tuned Hybrid Transformer Demucs (v4) at 9.0 dB SDR on the MUSDB HQ test set, against 5.9 for Spleeter. Spleeter wins on raw speed: Deezer says it splits to 4 stems '100x faster than real-time' on a GPU. If you would rather skip Python entirely, Veena splits any track into four stems in your browser, free, and lines them up on your timeline.

Which Demucs model is best?

For most songs, htdemucs_ft: the fine-tuned Hybrid Transformer model, which the Demucs README says takes four times longer than the default htdemucs but 'might be a bit better'. Use htdemucs_6s only if you need guitar and piano as their own stems, since its README warns the piano source 'is not working great'. Veena's built-in splitter gives you four stems with no model to choose.

Is there a Demucs v5?

No. On 16 November 2026 the newest release is Demucs 4.1.0, published on PyPI on 11 July 2026, and it ships the same v4 Hybrid Transformer models with lighter packaging and its weights on Hugging Face. 'Demucs v4' and 'htdemucs v4' name the same model family. Veena's splitter needs no version at all: right-click a clip in your browser and the stems land on your timeline.

What is the best Demucs alternative?

Veena, if you want stems you can keep producing with. Demucs hands you four audio files in a folder after a Python setup; Veena splits any track you upload into vocals, drums, bass and other in about 45 to 120 seconds, in your browser, free on every plan. The stems land on new tracks in your project, where CoProducer can build a new beat under your vocal.

Can I run stem separation models myself?

Yes. Demucs and Spleeter are open source and run locally with Python; Demucs works on CPU and much faster with a GPU. The maintained copy of Demucs now lives in its author's fork, whose README says to 'expect slow replies and no new feature for now'. For no setup at all, Veena separates a full song in about 45 to 120 seconds in your browser, free on every plan.

Meet your personal music producer

Free, in your browser. Describe the song you want and CoProducer builds it with you.

Make music free