MuScriptor — Public project availability¶

Development line: project:muscriptor · thread public-project-availability
Last event: 2026-07-22 · 1 dated since 2026-07-22 · Researched: 2026-09-05 · confidence: high

What it is¶

MuScriptor is an open-weight music transcription model from Mirelo and Kyutai. It turns multi-instrument recordings into MIDI note events.

• Checkpoints: small (100M), medium (300M) and large (1.3B) parameter models.
• Runtime: Python API, CLI and local HTTP serving.
• Input: 16 kHz mono audio in 5-second chunks.
• Output: pitch, onset, offset and 36-group instrument labels.

Weights are CC-BY-NC and gated; predictions do not include velocity.

Development line¶

• 2026-07-22 — MuScriptor public project resources were available. On 2026-07-22, public links opened for the project site, repository, model page, interactive demo and installation. These links confirm project availability on that date. They do not fix a specific release version, model capability or technical change.

What changed¶

• 2026-07-22 — Project links opened the public MuScriptor ecosystem. The preceding v0.2.2 release on 2026-07-21 added strict instrument-constrained decoding, direct MIDI HTTP output, Apple Silicon FP16/MPS support, optimized decoding kernels and Windows GPU instructions.
• 2026-07-21 — GitHub v0.2.2 expanded the runnable package alongside demo and service workflows.
• 2026-07-10 — Release v0.2.1 enabled serving without building. Kyutai published the project announcement.
• 2026-07-09 — The paper introduced the open-weight multi-instrument model. Training combined synthetic pre-training, real-audio fine-tuning and RL post-training.
• 2026-06-30 — Public checkpoint collection created. This marks repository metadata, not a training date.

How to use this¶

We can evaluate MuScriptor through its public code, model, demo and installation entry points since 2026-07-22. Confirm exact capabilities and versions from primary sources before adoption.

1. Install from the Git repository; PyPI installation remains listed as forthcoming. — https://huggingface.co/MuScriptor/muscriptor-large
2. Choose small, medium or large. Run muscriptor transcribe --model large audio.wav -o out.mid or call TranscriptionModel.load_model() and transcribe_to_midi() from Python. — https://huggingface.co/MuScriptor/muscriptor-large
3. Upload audio to the hosted transcription Space for an interactive trial. For local integration, use the repository CLI or server implementation. — https://huggingface.co/spaces/hugging-apps/muscriptor-music-transcription

Best practices¶

• Use instrument conditioning when track instrument groups are known. The model card notes that conditioning improves scores and keeps assignments coherent across chunks. — https://huggingface.co/MuScriptor/muscriptor-large
• Treat output as an editable MIDI draft rather than a finished score, especially on dense, unusual or heavily processed audio. — https://huggingface.co/MuScriptor/muscriptor-large
• Use only recordings with confirmed rights. The model card explicitly prohibits unauthorized transcription. — https://huggingface.co/MuScriptor/muscriptor-large

Superseded by this¶

• 2026-07-21 — v0.2.2 supersedes v0.2.1 for users needing direct MIDI HTTP responses, strict instrument constraints or native Apple Silicon FP16 inference.
• 2026-07-10 — v0.2.1 superseded v0.2.0 for serving without a build step.

Still unknown¶

• The 2026-07-22 event does not identify which linked surface changed that day. GitHub dates v0.2.2 to 2026-07-21, so that update is reported as the nearest first-party addition rather than a 2026-07-22 release.
• The schema supplies no separate fields for event_findings or new_events, so dated findings remain under what_changed.
• The hosted demo was reachable, but its rendered page text showed no operational documentation.

Sources¶

Agent brief¶

• Subject: project:muscriptor, thread public-project-availability, 1 dated events 2026-07-22 → 2026-07-22.
• Practical note: From 2026-07-22, practitioners could evaluate MuScriptor through its public code, model, demo, and installation entry points; confirm exact capabilities and versions from the linked primary resources before adoption.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.