Demo: switching between Apple FoundationModels and CoreML on iPhone (PFMSwitcher)

[john-rocky]

Recording of Examples/PFMSwitcher running on iPhone — one chat surface, picker flips the resident backend.

What you're seeing

• Same LanguageModelSession.streamResponse(to:) call site for both sides — the chat code never branches on which backend is active.
• Picker swaps between Apple FoundationModels (iOS 26+, when the toolchain links it) and mlboydaisuke/lfm2.5-350m-coreml (iOS 18+ on the Apple Neural Engine).
• Header shows live resident-set size via mach_task_basic_info. On the swap, RSS drops to the placeholder baseline first, then rises as the next model pages in — the strict release-before-load policy in ModelManager.switchTo at work.
• UIApplication.didReceiveMemoryWarningNotification aggressively releases the active backend so the app survives jetsam.

Try it yourself

git clone https://github.com/john-rocky/PrivateFoundationModels
open PrivateFoundationModels/Examples/PFMSwitcher/PFMSwitcher.xcodeproj

Xcode 16+ resolves the local SPM dependency. Build to a real iPhone (iOS 18+). First time you pick LFM2.5 350M the app downloads ~810 MB into ~/Library/Application Support/PrivateFoundationModels/.

To enable the Apple FoundationModels row, link FoundationModels.framework on an iOS 26 SDK toolchain. The bridge file is #if canImport(FoundationModels)-gated so the app still builds and runs on iOS 18 without it.

Verification

Harness	Status
swift test (unit)	55 / 55
swift run pfm-verify (e2e)	10 / 10
swift run pfm-portability (Apple-FM call sites compile + run unchanged)	8 / 8
swift run pfm-deep (every Generable × Tool combination)	PASS 7 / MODEL 4 / FAIL 0

Raw logs in docs/.