A 125M piano autocomplete model runs 108 notes/second on an iPhone 15, and DPO nearly tripled its preference score.

Decoder-only transformer with RMSNorm and rotary embeddings, using note-level tokenization NOTE(pitch, delta_onset, duration, velocity) so one forward pass advances the music by a complete note rather than an event fragment. Trained on a few hundred thousand MIDI files, ~300M note events. DPO moved pairwise preference from 24.55% to 69.05%, judged by Gemini 3.5 Flash. simedw.com 558 points on HN. Aggressively cleaned classical data beat larger noisier sets, which is the transferable lesson.