Show HN: I trained a 125M model to autocomplete piano on-device

A developer trained a 125M-parameter transformer model to autocomplete piano performances in real time, achieving approximately 108 notes/sec on an iPhone 15. Key improvements stemmed from optimizing the MIDI representation, aggressive data cleaning, and implementing DPO post-training. Pairwise evaluation using Gemini 3.5 Flash helped build a preference dataset for DPO, by asking it to compare continuations rather than assign absolute scores, and mirroring comparisons to reduce position bias.