Kyutai and the Mirelo team have released MuScriptor, an open-weight decoder-only Transformer for multi-instrument music transcription to MIDI. The model reads a mel-spectrogram of a short audio segment and autoregressively predicts MIDI-like tokens, treating transcription as a language-modeling task. Three weight variants are available on Hugging Face: small (103M), medium (307M, default), and large (1.4B). The inference code uses the MIT license, while the weights use CC BY-NC 4.0, restricting commercial use. Training proceeds through three stages: pre-training on D_Synth (roughly 1.45M MIDI files), fine-tuning on D_Real (an internal set of 170,000 recordings totaling more than 11,000 hours), and reinforcement learning post-training on D_RL (300 manually verified tracks). Evaluation on D_Test (372 held-out tracks) using the large model showed progressive improvement: the final model achieved an Onset F1 of 60.4, Frame F1 of 73.3, Offset F1 of 49.0, Drums F1 of 50.2, and Multi F1 of 48.2, compared to the YourMT3+ baseline scores of 32.5, 45.5, 17.8,