← Retour au blog
tech 20 August 2026

Training a 125M Model to Autocomplete Piano On-device

Discover how Simon Edwardsson developed a 125 million parameter model to autocomplete piano performances in real-time on an iPhone. An innovation that could revolutionize music creation.

Article inspired by the original source
Show HN: I trained a 125M model to autocomplete piano on-device ↗ simedw.com

Introduction

Imagine having a personal accompanist that completes your piano pieces in real-time, directly from your phone. That's exactly what Simon Edwardsson achieved by training a 125-million parameter model to autocomplete piano performances through an app called RollTab. Inspired by GitHub Copilot but for music, this project took nearly a year of development and optimization.

The Magic of MIDI Files

To understand how this tool works, it's essential to grasp what a MIDI file is. Unlike traditional audio formats like MP3, a MIDI file encodes music as a series of events: a key is pressed at a certain pitch and velocity, a key is released, the sustain pedal changes state, etc. These events are organized into multiple tracks, allowing for rich musical diversity.

Tokenizing Musical Events

One of the major challenges in developing this model was converting MIDI events into a discrete sequence that the model could read and predict. Instead of creating a token for every unique event, Simon opted for a factored approach using a reduced grammar. For instance, instead of having a vast combination of NOTE_ON and NOTE_OFF for each pitch and velocity, he structured the tokens as [NOTE_ON, PITCH, VELOCITY], significantly reducing the vocabulary size and easing the model's learning process.

Training the Model

The model, based on a 125M parameter transformer, was trained on cleaned and optimized MIDI data. The aggressiveness in data cleaning and selecting the right MIDI representation were critical to improving the model's performance. Additionally, a technique called DPO (Diverse Penalization Optimization) was added post-training to further refine the prediction quality.

Real-time Performance

One of the most impressive aspects of this project is its ability to run in real-time on an iPhone 15, producing around 108 notes per second. This opens up incredible possibilities for musicians, whether amateur or professional, to collaborate with AI and explore new creative dimensions.

Conclusion

The ability of an AI model to autocomplete musical performances in real-time on a mobile device marks a significant advancement in human-machine interaction. It could transform how we create and play music, making composition accessible to everyone. If you're curious about how this technology can be applied to your project, let's discuss it in 15 minutes.

Let's discuss your project in 15 minutes.

Feel free to contact us to delve deeper into this fascinating subject and explore how this technology can be integrated into your musical environment.

MIDI piano autocomplete AI music innovation
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call