Developer reports faster language model inference on a 2013 iPad Air

The smolmaxxing project reports running a 1.7B-parameter language model at 23.7-25.8 tokens per second on a jailbroken 2013 iPad Air, up from 2.75 with stock llama.cpp using the same model file. The prototype runs entirely on the tablet's CPU. Custom computation routines, faster output processing and changes to attention and token selection drive the improvement. A separate system loads selected model experts from flash storage, allowing an 8B mixture-of-experts model to run at 3.5 tokens per second. The iPad has 1 GB of RAM, with a practical process memory ceiling of about 700 MB. The developer cautions that the 1-bit Bonsai model still gets facts wrong and that the jailbreak does not survive a reboot.