Blankline Research2026.10.19 · 8m read
We did not set out to make music. Dropstone is a coding agent, and the model on its Fast tier, DeepSeek V4.1 Flash, is the smallest and cheapest one we ship. We asked it, from the CLI, to compose a short piece for full orchestra and deliver it in 7.1 surround, mostly to see where it would break.
It did not break. It came back with Nocturne in D minor: 282 notes for 15 instruments, 16 bars at 64 beats per minute, every section of the orchestra assigned to its own speaker the way a film-score mixer would assign it, and a 24-bit, 48 kHz, eight-channel lossless master. The whole run cost under $1.
The film above was cut to that master. YouTube re-encodes audio and plays stereo to almost everyone, so play the lossless master in the player just below the video, or download it and play it on a 7.1 system. Stereo is the least interesting way to hear this piece.
We Expected a Sketch #
A language model writing a tune is not new. What we expected was what we usually see: a melody, some chords under it, a part that sounds like a demo.
Orchestration is a different skill. It means knowing that a flute and a tuba live in different ranges, that violins carry a melody further than horns do, that a snare roll can pull a whole orchestra toward a resolution. Mixing for 7.1 is another skill again, and most human composers hand it to an engineer.
Dropstone did all three in one run. It wrote the notes, chose which of fifteen instruments plays each one, and then wrote its own routing sheet: which speaker every section comes from, at the angles the ITU-R BS.775 standard sets for a 7.1 room.
That is the part we did not see coming. Not that a model could write a tune, but that the cheapest one we run would think about where the listener is sitting.
Every one written by the model
Strings, woodwinds, brass and percussion
Eight speakers, a section on each
16 bars in D minor
Lossless, the same as the studio master
On the smallest model Dropstone ships
How It Learned to Write for Orchestra #
Dropstone's Fast tier is a general text model, not a music model. So it learned the way a developer would: from material that is freely available online. It studied orchestral MIDI files and notes on music theory and orchestration, and kept what it learned in its memory.
MIDI is a natural fit for a model that reads code all day. A MIDI file does not store sound. It stores instructions: which instrument, which note, how hard, when it starts and when it stops. To Dropstone, an orchestral MIDI file is structured data with a clear grammar, and a score is a program the orchestra runs.
Reading real scores this way teaches what a textbook states but only a score shows: the range each instrument is comfortable in, how sections double a line so it carries, how long a composer holds the brass back before letting it in.
You can hear that in Nocturne. The violins wait two bars before they enter. The orchestra builds for ten bars before the tutti. And the last chord is held and allowed to decay rather than cut off.
The Piece, Bar by Bar #
Nocturne is 16 bars at 64 beats per minute, which makes every bar exactly 3.75 seconds. It is slow enough for a nocturne, and it gives the film a clean grid: every cut lands on a bar line.
The shape is one long crescendo. For the first two bars the low strings, woodwinds, brass and percussion are already playing while the front speakers stay silent: the violins are being held back. They enter in bar 3 and reach full voice by bar 5. From there the whole room builds, through a snare roll on C sharp, the leading tone, in bar 12, to the full orchestra landing on D, the tonic, in bar 13. That is the loudest bar of the piece. Three more bars bring it down to a final held chord.
The curve below is not drawn by hand. It is the overall level of the master, measured from the audio.
Mixing Like an Engineer #
A conventional orchestral recording in surround keeps the orchestra in front of you and uses the rear speakers for the hall: reverb, air, applause. Dropstone did the opposite. It put the listener where the conductor stands and arranged the orchestra around them.
- Front left and right: first and second violins, carrying the melody, where your eyes and ears expect it.
- Centre: clarinets, horns and trumpets, the voices that anchor the middle of the texture.
- Side left: flutes, piccolo and oboes.
- Side right: timpani, snare, cymbals and bass drum.
- Rear left: violas, cellos and bassoons.
- Rear right: double basses, trombones and tuba.
- LFE: the lowest octave of the cellos, basses, tuba and timpani.
There is engineering logic in that layout. Woodwinds and percussion sit on opposite sides, so the brightest attacks in the orchestra do not mask each other. The low strings and low brass are split across the two rear speakers rather than piled into one corner, which keeps the bottom of the orchestra wide. And the sub carries only the instruments that actually reach that low, not a copy of everything.
The speakers sit at the standard ITU-R BS.775 angles: 30 degrees either side of centre for the front pair, 100 for the sides and 145 for the rear.
Eight Speakers, Eight Different Parts #
The easiest way to fake surround is to upmix: take a stereo track and spread it across eight speakers. It sounds big and says nothing, because every channel rises and falls together.
We measured each channel of the master separately, thirty times a second, to check that this is not what happened. It is not. The front pair is dark for the first two bars while the sides and rear play. Percussion and the sub peak in bar 10. The woodwinds, violas and cellos peak in bar 12, a bar before the tutti. Every speaker has its own part, its own rhythm and its own climax, which is what discrete 7.1 means.
Bar 6: Hearing the Difference #
The film makes the case in a single bar. In bar 6 the picture folds all eight orbs down to two, and the film's soundtrack does the same: the whole orchestra squeezed into the front left and right speakers, exactly as a stereo downmix would play it. In bar 7 it opens back out to eight.
On a 7.1 system the difference is physical. In stereo, the orchestra is a picture in front of you. In 7.1, you are standing inside it.
Under $1, on the Smallest Model #
Dropstone 1.8 runs three tiers. Fast, the one that wrote Nocturne, runs DeepSeek V4.1 Flash: the smallest, fastest and cheapest model in the lineup. We did not reach for Heavy, and we did not use a music model. The entire run, from studying the references to delivering the lossless master, cost under $1.
That matters more than the music does. Orchestral writing and a surround mix have always been studio work: a composer, an orchestrator, an engineer, a room full of speakers and a budget to match. Now anyone with the Dropstone CLI can describe what they want and get back a score, a routing sheet and an eight-channel master, for less than the price of a coffee.
What Is in the Film, and What Is in the File #
The lossless file below the video is the master exactly as Dropstone delivered it: the orchestra and nothing else.
The film's soundtrack adds a few cues for the picture. While each speaker is named on screen in bars 3 to 5, its sections are lifted a few decibels so you can hear where it is. Bar 6 is folded to stereo, as described above. A low sub drop sits under the tutti, a shower of small sparks travels from the rear to the front as the logo forms, and a bell and a low D ring under the end card. Our team added those for the film. Every note of music, in both versions, is Dropstone's.
How to Listen in 7.1 #
- Download the lossless master from the player under the video: FLAC, 24-bit, 48 kHz, eight channels, 34 MB.
- Play it in VLC, mpv, foobar2000 or Kodi with the output set to 7.1, or send it over HDMI to an AV receiver.
- Channel order is the standard one for FLAC: front left, front right, centre, LFE, rear left, rear right, side left, side right. Any player that reads the file's channel layout will route it correctly.
- On headphones or a laptop, your player will fold it to stereo. It still sounds good. It is just not the point.
Make Your Own #
Everything in this post was done from the Dropstone CLI on the Fast tier. Install it on macOS, Linux or Windows, then describe the piece you want, the instruments you want in it and the room it will play in.
brew tap blankline-org/dropstone-cli
brew install dropstone
curl -fsSL https://dropstone.io/install | bash
irm https://www.dropstone.io/install.ps1 | iex
Frequently asked questions #
Can Dropstone compose music?
Yes. Dropstone is a coding agent, but music written as MIDI is structured data. From the CLI, Dropstone 1.8 composed, orchestrated and mixed Nocturne in D minor: 282 notes for 15 instruments, delivered as a 7.1 surround master.
Which model composed Nocturne?
Dropstone 1.8 Fast, which runs DeepSeek V4.1 Flash. It is the smallest and lowest-cost model in the Dropstone lineup.
How much did it cost?
Under $1 for the entire run.
Can I hear it in 7.1 on YouTube?
Not reliably. YouTube re-encodes audio to a lossy format and plays a stereo mix in most browsers. For true 7.1, download the lossless FLAC master from this page and play it on a 7.1 system.
What format is the download?
FLAC, 24-bit, 48 kHz, 7.1 (eight discrete channels), 68.7 seconds, 34 MB.
“We asked a coding agent for a sketch. It handed back a score, a routing sheet and a mix.”
Conclusion #
We usually measure Dropstone on code: benchmarks, agent sessions, pull requests. Nocturne was not on any roadmap. It is a reminder that a capable model with memory and a terminal is not limited to the job it was named for.
Composition, orchestration and a discrete surround mix are crafts that take people years. The smallest model we ship did all three, in one run, for under $1. Watch the film, download the master, and if you have eight speakers, turn it up.
Then make something of your own and send it to us. We would like to hear it.