Show HN: I trained a 125M model to autocomplete piano on-device
Summary
Simon Edwardsson trained a 125M-parameter transformer to autocomplete piano performances on-device, achieving about 108 notes per second on an iPhone. He emphasizes MIDI representation, aggressive data cleaning, augmentation, and Direct Preference Optimization (DPO) to improve continuations, plus on-device deployment with Core ML. The post covers model design, evaluation, and practical trade-offs for running medium-sized AI models on mobile hardware.