Apple Silicon is fast enough now that local-first AI is no longer a science project. With the right quantization and a careful prompt, a 7B model on an M-series machine can hit useful latency for chat, summarization, and lightweight agent work.

The tradeoffs are real. You give up the ceiling of frontier models. You spend more time on context management. But you get privacy, offline capability, and no API bill at the end of the month.

For Hank, that calculus made sense. The app lives on the user's machine. Their files, their context, their compute. No round trip to a cloud.