Sadman Habib

OffLlama is an iPhone assistant that keeps working when there's no signal. I'm a computer science student at Hunter College, and my New York subway commute loses connection constantly. That's exactly when assistants go silent. And the ones that can act on my phone rarely show me what they're about to do. OffLlama runs Llama-Reflex-1B directly on the phone as its System 1. It's a Llama 3.2 1B model I distilled from Muse Glimmer 30B. In one forward pass, it scores a small set of options: answer locally, use a tool, ask a clarifying question, or escalate. It trusts its own choice only at 0.90 confidence or higher. Below that, it takes a safe fallback instead of guessing. Supported requests become native iOS actions, including Reminders and Calendar, Messages and Mail compose sheets, and Apple TV search. Each action is drafted for review before anything changes. When I'm online, harder questions can go to a Muse model as System 2, after on-device text screening. On my iPhone 17 Pro, it answers in Airplane Mode, saves reviewed reminders, and opens Apple TV searches. On 120 hand-labeled decisions, it picks the right option 98 times, up from 89 before distillation. At the 0.90 threshold, it accepted 44 decisions and got 41 right. The repository includes the model, benchmarks, and the experiments that failed.