Upload an audio file and Jukeblocks separates it into stems, then builds a project with each stem on its own track.
The song is split into 4 stems: vocals, drums, bass and everything else.
Pick your DAW in the dropdown. Only that format is exported, so choosing Ableton gives you an .als and nothing else.
You get a zip with the project file and the audio it references. Keep them together: the project looks for the audio in the same folder.

This is a beta feature, available on Pro plans while it is being developed. Separation quality depends on the source material, and dense or heavily processed mixes bleed more between stems.

Most stem splitters hand you a folder of wav files. This one hands you a project, tempo detected and every part already on its own track, so you start working instead of importing.

Remixing a song
A remix starts with the parts, not the mixdown. Upload the original and you get the vocal, drums, bass and the rest as separate tracks at the song's tempo, so the first thing you do is choose what to keep rather than fight with alignment. Each stem is chopped into 4 bar loops instead of one long file, so you can move the chorus vocal, repeat a bar or delete a section without slicing anything yourself. The melodic parts also come through as MIDI, which is the quickest way to re-play a bassline on your own sound instead of reusing the recording.

Bootlegs
A bootleg is someone else's vocal over your own production, so the part you need is the vocal stem. Build underneath it and use the detected tempo to line the two up. The thing that catches people out is ambience: reverb and delay printed on the original vocal come along with it, so a dry backing track can sound oddly wet underneath. Leaving space where the original ambience sits usually works better than trying to remove it.

Mashups
Run both songs through and you get two projects, each with its parts separated and its own tempo detected. From there the job is mostly tempo: decide which song sets the pace and stretch the other to match. The detected tempos tell you how far apart they are before you start, which is usually the deciding factor, since a few BPM stretches transparently and twenty does not. Key matters too, and shifting a vocal more than a couple of semitones starts to show.

Stem separation on its own
If you do not want a project at all, the zip contains every stem as a standard wav file: vocals, drums, bass and everything else. The stems line up with each other and with the original, so layering them back together reproduces the track.

Instrumentals and acapellas
Removing the vocal and isolating the vocal are the same job done twice, and both come out of the same split: mute the vocal track for an instrumental, solo it for an acapella. How clean the result is depends almost entirely on the source. A forward vocal in a sparse arrangement separates well, while heavy processing is the hardest case, since a vocal drenched in reverb leaves some of that tail behind in the instrumental.

Learning a track
Solo the bass to hear exactly what is played, or the drums to hear the pattern without the rest on top. The melodic parts come with a MIDI transcription on their own track, so you can watch the notes in the piano roll while the audio plays. It is most accurate on clear monophonic parts like a bassline, and least accurate on dense chordal material.

The tempo is detected from the audio, rounded to a whole number, and the project is set to it.
Stems are chopped into 4 bar loops so you can move sections around instead of dragging one long file.
Songs that do not start exactly at 00:00 are handled: the leading silence is measured and the clips are aligned to it.
Live or unquantised recordings drift against a fixed grid, so expect some manual nudging on material that was never played to a click.

The melodic stems are also transcribed to MIDI and placed on their own tracks, each loaded with an instrument so they play straight away.
Those MIDI tracks are muted by default, because the audio is what you hear first. Unmute one to play the transcription instead.
Drums are split out as audio but not transcribed: pitch detection on unpitched material produces noise.
You can pick the instrument in More Options. Transcription is an estimate, so treat the MIDI as a starting point for a re-play rather than an exact copy.

Under More Options you can choose what the transcribed MIDI tracks are loaded with.

Random Presets (default): a stock synth preset from your own DAW, picked to suit each stem. Ableton gets Analog, Operator, Collision or Electric; FL Studio gets Harmor or Sytrus. Nothing is sampled, so the tracks always make sound and the project stays small. The sound is a starting point, not an imitation of the original.

Single Sample: one note is captured from the stem and loaded into a sampler, Simpler in Ableton and the Sampler channel in FL Studio. You hear the actual recorded sound rather than a synth, but a single sample is stretched across the keyboard, so notes far from the captured pitch start to sound unnatural.

Multi Sampler: one sample per note is captured, using the longest example of each pitch found in the stem, and each is mapped to its own key zone. Every note then plays a real recording of that note instead of a stretched one. This applies to bass and vocals, where the part is one note at a time; other stems fall back to Single Sample, because a slice of a chord makes a poor instrument. Ableton only, and it needs Live Suite for the Sampler device.

All of these load muted, so the audio is what you hear first. Unmute a MIDI track to play its instrument instead.

Separation is the slow part and runs on the server, so a full song takes a few minutes.
The job runs in the background: the progress bar tracks it and the download starts on its own when it is ready.