MuScriptor is a multi-instrument music transcription model — audio in, MIDI out — built by Kyutai and Mirelo, and it has picked up 1,059 stars and 134 forks since the repository opened on 2 July 2026. The project calls itself "the most accurate open-source transcription model," a claim it backs with an accompanying paper, MuScriptor: An Open Model for Multi-Instrument Music Transcription (arXiv:2607.08168), authored by Simon Rouard, Michael Krause, Axel Roebel, Carl-Johann Simon-Gabriel, and Alexandre Défossez — the same researcher behind Demucs and Moshi.

The architecture is deliberately plain: a transformer decoder only, published in three sizes. small is 103M parameters (14 layers, width 768), medium is 307M (24 layers, width 1024) and is the default speed/accuracy trade-off, and large is 1.4B (48 layers, width 1536). The team's guidance is that small is the realistic pick on a CPU-only machine while large really wants a GPU. On Apple Silicon the model moves to Metal (MPS) without configuration.

What separates this from a research drop is the delivery. Because the package is on PyPI, uvx muscriptor transcribe path/to/audio.wav runs a transcription with no repository clone at all, and uvx muscriptor serve starts the same web UI that Kyutai hosts at muscriptor.kyutai.org, complete with a live piano roll. The CLI's --format sheets flag goes a step further than MIDI and engraves actual readable notation: the output directory gets score.mid, score.musicxml, a full_score.pdf with every instrument on one system, then one PDF per instrument plus a tablature PDF for fretted ones — electric guitar, electric bass, drum kit, each split out. That path needs MuseScore 4 or newer installed separately.

Platform handling is unusually careful for a model repo. Windows users need --torch-backend=cu128 to get off the CPU default; Intel Macs need --python 3.12 pinned, because PyTorch stopped shipping x86_64 wheels after torch 2.2.2. Both quirks are documented in a table rather than left in the issue tracker.

Two caveats matter before adopting it. First, and most important, the code and the weights are not under the same license: the repository is MIT, but the model weights on HuggingFace are CC BY-NC 4.0 — non-commercial only. Anyone planning a paid product around this needs to resolve licensing with Kyutai first, and the distinction is easy to miss when GitHub's sidebar simply reads "MIT." Second, the weights are gated: you have to log into HuggingFace and accept the license on each model page before uvx can pull anything, so there is no fully anonymous install path, and CI or air-gapped setups need an HF_TOKEN provisioned in advance.