apple-foundation-models

Apple Foundation Models (On-Device AI)

Build private, offline, no-cost AI features on Apple's on-device foundation model (iOS 26 / macOS 26 / Apple Intelligence), via import FoundationModels. The model runs on-device: zero server cost, works offline, data never leaves the device.

Verify signatures as you go. This framework is new and evolving. Use the apple-docs MCP (check_availability for OS/version gating, get_symbol for exact API shapes, list_framework FoundationModels) before committing to a signature. The patterns below are stable; treat specific initializer/parameter names as "confirm against live docs," in the spirit of ios26-api-reference.

---

Right-sizing: what this model is (and isn't)

The on-device model is a small (~3B-class) language model, not a frontier chatbot.

Great at: summarization, classification, tagging, extraction, rewriting, short-form generation, structured output from unstructured text, semantic routing.

Not for: authoritative world knowledge, math/code reasoning at scale, long documents beyond the context window, anything where a confident hallucination is unacceptable. For those, call a server model — don't force the on-device model past its weight class.

If a task needs world facts, ground it: pass the facts in the prompt (retrieval), don't expect the model to know them.

---

1. Gate on availability — always

The model is absent on ineligible devices, when Apple Intelligence is off, or while assets download. Check before showing any AI UI.

import FoundationModels

let model = SystemLanguageModel.default

switch model.availability {
case .available:
    // show the feature
case .unavailable(let reason):
    // .deviceNotEligible, .appleIntelligenceNotEnabled, .modelNotReady — degrade gracefully
    break
}

Never assume availability. Provide a non-AI fallback path for every AI feature (older devices, EU/region/enterprise restrictions, model still downloading).

---