Simon Edwardsson · 2026-08-20 · notable
RollTab — a 125M piano model that autocompletes on an iPhone
RollTab is a free iOS app that autocompletes piano playing from a MIDI keyboard. A 125M-parameter decoder-only transformer runs on the phone at about 108 notes per second on an iPhone 15, trained on roughly 300 million note events.
RollTab plugs a MIDI keyboard into an iPhone and finishes the phrase you started, entirely on-device.
Key specs
| Parameters | 125M |
|---|---|
| On device speed | ~108 notes/sec on iPhone 15 |
What is it?
RollTab turns a MIDI piano and an iPhone into an autocomplete pair: you play a few bars, and a 125M-parameter model continues the piece in real time. Simon Edwardsson published the write-up on 20 August 2026 and shipped the app free on the App Store. Everything runs locally through Core ML, so there is no server call between a keypress and the continuation.
How does it work?
The tokenizer is where most of the speed comes from. Instead of emitting separate TIME_SHIFT events, Edwardsson encodes each note as a single compound NOTE(pitch, delta_onset, duration, velocity) token, so the transformer advances the music one whole note per forward pass — roughly 5x fewer autoregressive passes per note. The network itself is a standard decoder-only stack with RMSNorm, rotary position embeddings, causal self-attention and SwiGLU blocks. DPO post-training raised the preference score from 24.55% to 69.05%.
Why does it matter?
Small on-device models are usually shown off on text; this is a working music example with latency low enough to play against. Training used a few hundred thousand public-domain classical MIDI files, about 300 million note events, and heavy cleaning and deduplication — Edwardsson credits the data work and the token format, not scale, for most of the gain. He rates the output at roughly GPT-2 level, with rhythm and long-range structure still weak.
Who is it for?
musicians with a MIDI keyboard, on-device ML tinkerers
Try it
https://apps.apple.com/gb/app/rolltab/id6799679058