TECH Signal 501
125M transformer autocompletes piano at 108 notes/sec on iPhone
↗
A developer trained a 125M-parameter transformer that autocompletes piano performances in real-time on an iPhone, achieving approximately 108 notes per second.
This demonstrates that small transformer models can produce musically coherent output on mobile hardware without cloud dependency. The key insight is that representation design, specifically encoding each note as a single token with explicit duration, eliminates drift problems and reduces computational cost per note.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
The final MIDI representation encodes each note as one token with pitch, delta_onset, duration, and velocity, requiring one transformer pass per note instead of four.
The model runs at approximately 108 notes/sec on an iPhone 15, exceeding the rate needed for live piano performance.
The biggest improvements came from the MIDI representation choice, aggressive training data cleaning, and DPO post-training.
THE CLUSTER