Weekend hack runs real-time music AI on iPhone 12 — no GPU needed
Google DeepMind released Magenta Realtime 2, an open-source AI that generates music live, but said iPhones were not supported. One developer proved otherwise, running it for 10 straight minutes on a 2020 iPhone 12 Pro without overheating. The trick was splitting the model across different parts of the phone's chip and using the built-in NPU instead of the GPU.
Magenta Realtime 2 is DeepMind's newly released open-source model that creates music in real time. The official announcement said it works on Mac but not iPhone. Rather than accept that, the developer spent a weekend splitting the AI model into five separate pieces and assigning each piece to a different part of the iPhone's chip — CPU, NPU, and others — so no single component gets overwhelmed.
The key insight is that iPhones have a dedicated NPU (neural processing unit) that handles AI tasks far more efficiently than the GPU. Using the GPU for sustained AI work on a fanless phone risks overheating and damage — something the developer had experienced firsthand before. This proof-of-concept shows that powerful on-device, real-time AI audio generation on older iPhones is achievable without a server connection.
Key points
- DeepMind's Magenta Realtime 2 music AI now runs on iPhone despite official claims it couldn't
- Tested on a 2020 iPhone 12 Pro — 10 minutes continuous, no overheating
- Avoided the GPU entirely by routing work through the iPhone's built-in NPU
- Model was split into 5 pieces distributed across different parts of the chip
- Opens the door to offline, real-time AI music apps on existing iPhones
Quick term guide
- Magenta Realtime 2
- An open-source AI model from Google DeepMind that generates music in real time.
- open-source
- Software whose code is shared publicly so others can inspect, use, or change it.
- open-source model
- An AI model whose code and weights are freely available for anyone to download and use.
- AI model
- A program that can understand prompts and produce text, code, or answers.
- AI Mode
- A Google Search feature that uses AI to answer longer, more detailed questions.
- proof-of-concept
- A small demo built to show that an idea actually works, not a finished product.
- on-device
- Running an AI model directly on your phone or computer instead of sending data to a remote server
- routing
- Automatically deciding which AI model handles a request based on how complex or simple it looks.