A developer has trained a 125M-parameter transformer model to autocomplete piano performances in real time – directly on an iPhone 15, at a speed of about 108 notes per second. The idea is inspired by GitHub Copilot or Tabnine, but instead of prompting with code, you play a few notes on a MIDI piano, and the model continues where you left off. The entire process runs locally on the device, which opens up new possibilities for creative software and music production.
For decision-makers in the music and creative software industry, this is a signal that AI-driven creative tools no longer require large cloud resources. Running the model on-device means lower latency, better privacy, and the ability to work offline – factors that can be crucial for products that need to function in real time, such as digital audio workstations or live performance.
Although this is a hobby project and not a commercial product, it demonstrates that even relatively small models can deliver impressive results when they are specialized. This could inspire new business models where AI assistance becomes a standard feature in existing music software, rather than a separate cloud-based offering.
For those considering building or investing in creative AI tools, it is worth following developments in on-device models. The technical details about the model, training, and Core ML integration are described in the developer's blog post, and there is also a free app to try out the functionality.
Conclusion: On-device AI models like this can make creative tools faster and more private, which could change the competitive landscape for music software and other real-time applications.
