Apple engineers' recap for local AI builders: skip AFM for coding, one-line model swapping
rxwei · x · 2026-11-09
A full recap of this week's Apple Intelligence Group Lab session with engineers from Apple's Foundation Models, Core AI, Siri and App Intents teams, aimed at local AI builders.
- One API over every model: Apple's new app framework can swap between the on-device model, cloud model, or an open-weight model converted via Core AI with a one-liner — external models get the same tool calling, structured output, and Instruments profiling as Apple's.
- Don't use AFM for coding: Apple engineers themselves say bring your own model via Core AI; one engineer mixes MLX and Core AI models across multiple coding harnesses with good results.
- Agent Siri needs the cloud: Siri's tool intents run only on Private Cloud Compute; offline, your app's tools sit idle.
- fm serve: macOS 27 ships a local Chat Completions server for the on-device model.
- Context windows: AFM Core (3B, on-device) 4K–8K; AFM Core Advanced (20B, on-device) 8K; AFM cloud 32K; your own model via Core AI (e.g., Gemma 4 E2B/E4B) 100K. Only one on-device model per device.
More from coding & agent
- Connecting Grok Bot to Claude Code creates an AI loop that red-teams business plans — adamamcbride · 2026-11-09
- levelsio says Claude Code can design your furniture cheaper than brands; critics do the math — rms80 · 2026-11-09
- Daniel Mac shares 7 design principles for his Claude Duet + Projects + Code personal agent — daniel_mac8 · 2026-11-09
- Ad agency auto-builds a weekly research pack from 8 data sources, saving strategists hours — alexgoughcooper · 2026-11-09
- Should you let AI agents pay for you? Keep the final approval, argues dev — Efistoffeles · 2026-11-09
- Burkov: Codex on phone should stream desktop browser since localhost dev is unreachable from mobile — burkov · 2026-11-09