On-Device AI and Apple Apps
A model on the device costs nothing per call, works offline and keeps the user’s data where it is,
which is the architecture privacy-first asks for. The price is a small model and a hard platform.
These pages are what I measured getting real features to run that way, mostly inside
project-life2film. Shrinking a model for the device is the solo:model-shrink skill.
Models on the device
- apple-on-device-ai: one on-device layer for all my iOS apps, built on Apple’s Foundation Models
- apple-on-device-agent-sgr: Apple’s 3B model planned tool calls right 11 times in 24, and what fixed it
- on-device-beat-tracking-downbeats: a 20M-parameter net found downbeats where the DSP detector missed nine in ten
- shot-detection-phone-footage-no-cuts: sixteen shot detectors on two days of phone footage
- needle-tiny-tool-model: a 45M model in 14 MB that only picks tools and fills arguments
- bonsai-ternary-llm: a 500M ternary model running in the browser
The platform around them
- swift6-callback-isolation: the Swift 6 crash when a framework callback inherits main-actor isolation
- superduper-analytics: every app and site measured on one self-hosted counter, no cookies
- posthog-cross-platform: the analytics setup it replaced, and why
- kmp-sync-strategy: offline-first sync between Android and iOS