Shipping a 1 GB language model inside an iPhone app
Download sizes, memory pressure and why the simulator cannot help you here.
The model behind AI greetings is about 1 GB in its 4-bit form. We do not bundle it with the app. Cake downloads it the first time you use the feature and keeps it in the device cache.
Memory
On iPhone, the GPU cache is the biggest lever against memory pressure. We cap it tightly and release the model when you leave the greetings screen.
The simulator has no Metal GPU for this workload, so Cake shows a clear message there instead of crashing.