musicar

What musicar sounds like

musicar is an open model that writes music as MIDI. It composes from a short text prompt (English or Ukrainian) or continues a piece you give it, for any mix of 129 instruments. The 464M-parameter model was trained from scratch on one rented RTX 5090. Audio is rendered with a standard General MIDI soundfont, so the sound is plain; the notes are the model's.

How these were picked: by a scoring model (CLaMP 3), not by listening. For each style the clip is the top-scoring of 8 candidates (32 for the string piece), so these are the model's best cases per style. For an unpicked random draw next to other models, see the side-by-side page.

From a text prompt

Music from the prompt alone, no opening, released settings. First 45 s of each.

Prompts that did not work: an epic orchestral film score (strings, but no brass or timpani), a Ukrainian folk song with accordion (the model never produced an accordion), and a slow, sad Ukrainian piano melody (drums and bass crept in). Long, detailed prompts with key and tempo are its weak spot.

Continuing a real piece

The first seconds are a real opening, the rest (30 s) is musicar. Rank = where the scoring model puts musicar's continuation among 96 real ones when asked which is closest to the true continuation.

The openings are short excerpts (synthesised from MIDI) of existing pieces from the GigaMIDI and Aria-MIDI research datasets, here only to show the model; the rights to the music stay with their authors. To have an excerpt removed, open an issue in the repository.