Decomposer: Decompile Symbolic Music to Programs

Researchers Yewon Kim, Apurva Gandhi, David Chung, Graham Neubig and Chris Donahue released Decomposer, an 8B-parameter model that decompiles symbolic music (MIDI) into executable, editable Strudel programs, detailed in arXiv paper 2607.01849. The release ships inference code with two backends — generate_transformers.py and generate_vllm.py — running on a GPU with 24GB+ VRAM and CUDA 12.4+, sampling 5 candidate programs by default at temperature 1.0 with a 4096-token generation cap. The model was mostly trained on clips under 30 seconds, so the authors recommend decompiling 15–30 second windows and filtering dense instrument selections that exceed the input limit.

📄 Paper https://arxiv.org/abs/2607.01849 | 🌐 Website https://yewon-kim.com/decomposer | 🎹 Live Demo https://huggingface.co/spaces/haiyewon/decomposer-demo | 🤗 Model https://huggingface.co/haiyewon/Decomposer-Qwen3-8B Decomposer decompiles symbolic music MIDI into executable, editable Strudel https://strudel.cc programs. This repo contains the inference code for our post-trained model. Requires uv https://docs.astral.sh/uv/ A GPU with 24GB+ VRAM and CUDA 12.4+ is recommended uv sync transformers backend uv sync --extra vllm vLLM backend We provide two inference backends with the same command-line interface: - generate transformers.py — simpler setup, slower generation - generate vllm.py — faster generation, heavier setup | Parameter | Default | Description | |---|---|---| | --midi path | required | input MIDI file .mid | | --model | haiyewon/Decomposer-Qwen3-8B | HF model id or local path | | --n outputs | 5 | number of Strudel programs to sample | | --start sec | file start | start of the time window to decompile seconds | | --duration sec | whole file | length of the time window seconds | | --instruments | all | comma-separated instruments to keep: GM program numbers 0–127 , GM names e.g. acoustic bass , or drum | | --list instruments | None | print the instruments found in --midi path and exit | | --bpm | from file | tempo override | | --meter | from file | beats-per-measure override | | --temperature | 1.0 | sampling temperature | | --max new tokens | 4096 | generation length cap | | --output root | outputs/ | directory for generation sessions | | --log prefix | None | session directory prefix: outputs/<log prefix <midi name / ; defaults to timestamp if None | decompile a whole MIDI, sampling 5 candidate programs uv run generate transformers.py --midi path assets/examples/1.mid --n outputs 5 decompile a specific 15-second window of your own MIDI uv run generate vllm.py --midi path song.mid --start sec 30 --duration sec 15 see which instruments a file contains uv run generate transformers.py --midi path song.mid --list instruments decompile only selected instruments GM names, program numbers, or 'drum' uv run generate vllm.py --midi path song.mid --instruments drum,acoustic bass,4 Each run writes a session directory: outputs/20260719 123456 song/ ├── input repr.txt the serialized MIDI the model saw ├── output 1.js generated Strudel programs … ├── … └── args.json Paste any output .js into the Strudel REPL https://strudel.cc to play and edit it. - The model is mostly trained on clips under 30 seconds; a range of 15–30 seconds would work best. - Very dense selections many notes / many instruments can exceed the model's input limit; shorten the window or filter instruments. - Sampling uses temperature 1.0 the paper's evaluation setting ; we recommend sampling multiple outputs --n outputs 5 and picking your favorite. @article{kim2026decomposer, title = {Decomposer: Learning to Decompile Symbolic Music to Programs}, author = {Kim, Yewon and Gandhi, Apurva and Chung, David and Neubig, Graham and Donahue, Chris}, journal = {arXiv preprint arXiv:2607.01849}, year = {2026} }