A developer trained a 125M-parameter transformer that autocompletes piano performances in real time on an iPhone 15, running at ~108 notes/sec entirely on-device using Core ML. The concept mirrors tools like GitHub Copilot, but prompts the model by playing MIDI notes instead of code. The app is free to try, and the author offers to discuss training details and lessons learned.
Background
On-device AI inference has become a growing trend as mobile chips gain enough compute power to run smaller language and generative models locally. Core ML enables Apple devices to run ML models with low latency and no internet dependency.
- Source
- Hacker News (RSS)
- Published
- Aug 20, 2026 at 08:04 PM
- Score
- 6.0 / 10