gg2
Piano autocomplete model trained on-device
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
A developer trained a 125M-parameter transformer to autocomplete piano on-device at ~108 notes/sec on an iPhone.
The app is free, and the model runs locally, making AI-assisted music creation private.
Background
The developer, Simon Edwardsson, spent nearly a year experimenting with connecting a MIDI piano to a phone and having AI autocomplete songs. After 14 experiments, he trained a 125M-parameter transformer that runs entirely on-device, similar to GitHub Copilot but for piano. The biggest improvements came from finding the right MIDI representation, cleaning training data aggressively, and adding DPO post-training.
How it unfolded
- Aug 20, 2026Developer Simon Edwardsson published a blog post and Show HN on Hacker News announcing the 125M-parameter piano autocomplete model, running at ~108 notes/sec on an iPhone 15.
- Aug 20, 2026 12:00The Show HN post appeared on Hacker News, quickly gaining points and comments.
- Aug 20, 2026 18:00The story was picked up by aggregator sites like Zeli, summarizing the app's features.
Who’s saying what
- Developer
- Simon Edwardsson says the model works well enough to share, and he's happy to answer questions about the model, training, and Core ML.
- Public
- Hacker News readers showed interest, with the post receiving 17 points and 5 comments within hours.