24,265 Downloads. Its README Tells You When to Pay a Competitor.

TL;DR

StemDeck is a free, Apache-2.0 desktop app that splits a song into six stems (vocals, drums, bass, guitar, piano, other) without an account, an upload, or a subscription. Since its first release on May 4, 2026 it has shipped 72 releases and 24,265 installer downloads, and it currently sits at 2,562 stars and 246 forks. Two things make it worth your attention beyond "another Demucs wrapper." First, the model doing the actual work has not received a commit since April 2024 and its home repository is archived. Second, the README contains a comparison table where the author lists, in detail, every dimension on which the paid cloud competitors beat their own software. That second thing is rarer than the first.

What it does

Drop in an MP3, WAV, FLAC, OGG/Opus, MP4, or M4A file and StemDeck returns up to six isolated tracks, then loads them into a DAW-style multitrack mixer: per-stem volume faders, mute, solo, live VU meters, waveform zoom, loop regions, and a WAV export of whatever subset you selected. It also runs BPM and key detection through librosa and integrated loudness through pyloudnorm against the ITU-R BS.1770 standard, so you get the numbers a mastering engineer would ask for before you have opened anything else.

Source separation is the part worth explaining properly. You hand the model a stereo mixdown, which is a single waveform where every instrument has already been summed together, and it returns six waveforms that add back up to roughly the original. Think of it as being handed a baked cake and asked to return the eggs, the flour, and the sugar in separate bowls. The model has never seen the ingredients, only cakes, and it has learned what eggs tend to do to a sponge.

The practical payoff for a musician: you can pull the bass line out of a track to learn it, mute the vocal to build a backing track, or lift a drum loop, all without sending an unreleased master to somebody else's server.

The engine is a model nobody maintains

StemDeck defaults to htdemucs_6s, the six-source variant of Hybrid Transformer Demucs, published by Meta AI researchers in the paper Hybrid Transformers for Music Source Separation. The weights are about 170 MB and download on first run.

Here is the part that should make you sit up. The original facebookresearch/demucs repository, 10,359 stars and MIT licensed, is archived and read-only, with no code pushed since April 2024. Maintenance moved to a personal fork by one of the original authors, which has 3,101 stars and last saw a commit in July 2026. The state of the art in open music source separation has, for practical purposes, been frozen for over two years, and it is still good enough that a four-month-old hobby project can put it in front of 24,000 people.

And a caveat StemDeck inherits: the Demucs authors call the six-source model experimental, reporting "okay quality for guitar, but a lot of bleeding and artifacts for the piano source." That 9.0 dB headline number belongs to the fine-tuned four-source model. If your piano stem sounds like it wandered in from the next room, that is the model, not the app.

Nothing leaves the machine

The architecture is the whole product pitch. There is no server-side anything: a FastAPI backend runs on localhost, streams job progress over Server-Sent Events, shells out to FFmpeg and PyTorch, and the frontend is vanilla JavaScript with the Web Audio API and canvas waveforms, no framework and no build step. Torch device selection is automatic: CUDA on NVIDIA, MPS on Apple Silicon, CPU if neither is there.

Packaging is Tauri v2, which pays off in a way worth stealing. The macOS DMG is 7.8 MB. It downloads its own Python runtime, FFmpeg, and model weights on first launch instead of shipping them, so the thing you click on is small and the 500-plus MB of runtime arrives once and gets cached. The Windows and Linux bundles take the opposite approach and ship fat, at roughly 285 MB and 393 MB respectively. There is also a Docker image on GHCR, an Unraid Community Applications entry with GPU passthrough notes, and a plain uv sync path if you would rather just run the FastAPI server yourself.

The table that makes this a story

Most open-source READMEs are recruitment posters. StemDeck's contains a section titled "Honest Comparison" that spends a full table explaining where Moises and LALAL.AI are simply better: proprietary models that are "regularly updated, generally higher quality," up to ten stems instead of six, batch processing, iOS and Android apps, pitch shift, chord detection, lyrics, click track. Its own row for polish reads "Functional, hobby-grade UI." The closing line tells you outright that if you need speed, quality, mobile access, or deeper musician tooling, "the commercial products are worth the money."

The README also states the project "does not accept any money, sponsorship, or funding" from any of the makers and communities it links to, and then links a dozen luthiers, pedal builders, and subreddits anyway.

This is a positioning strategy, not just good manners. By naming the ceiling up front, the project defines the exact user it wants (someone learning a part, someone who will not upload an unreleased track) and stops burning issue-tracker time on people who arrived expecting a cloud product. Underpromising is cheaper than support.

What to actually check before you install

• Piano is the weak stem. Not an app bug. The upstream authors flagged it in the model card.
• There is a length cap. STEMDECK_MAX_DURATION_SEC defaults to 1200, so 20 minutes. Raise it if you are working on a DJ set.
• Jobs expire. Job directories are swept on a 24-hour TTL by default. Export what you want to keep.
• One job at a time. No batch queue, and the pending-job limit is 3 before the API returns a 503.
• CPU-only is slow. The Intel Mac and non-NVIDIA Windows builds work, but a transformer separating six sources across a full song is not a task you want on cores.
• The YouTube import is a legal question, not a technical one. It runs yt-dlp, and the project's own disclaimer says automated downloading may violate YouTube's terms and puts responsibility on you.

The optional lead-versus-backing vocal split is a second model entirely: UVR-MDX-NET Karaoke 2, run through python-audio-separator and trained as part of the Ultimate Vocal Remover project, which has 26,025 stars of its own. Both are MIT licensed. The entire stack a working musician needs for this task is now free, open, and largely finished.

Key Takeaways

• Frozen models are still shippable products. Demucs has not taken a commit at its original home since April 2024, and a 2026 app built on it has 24,265 downloads. Not every category needs a new state of the art each quarter.
• Local is a structural privacy claim. Cloud splitters require you to upload unreleased audio and trust a retention policy. StemDeck removes the network hop entirely, which is a different kind of promise.
• Publishing your own weaknesses is a positioning move. The "Honest Comparison" table names ten places the paid tools win, and in doing so tells the right user to stay and the wrong user to leave.
• Thin installer, fat first launch. The 7.8 MB Tauri DMG that fetches its runtime on first run is a pattern worth copying for any Python-plus-PyTorch desktop app.
• Check which model variant a wrapper defaults to. The 9.0 dB SDR number in the Demucs paper is the fine-tuned four-source model; the six-stem default is explicitly experimental.
• Windows still owns local creative AI. On the current release, Windows bundles outpulled macOS 1,049 to 627 and Linux by roughly nine to one.

Sources: StemDeck on GitHub, StemDeck releases, facebookresearch/demucs (archived), adefossez/demucs, Hybrid Transformers for Music Source Separation (arXiv:2211.08553), Ultimate Vocal Remover, python-audio-separator, Hacker News discussion