Three Apple language paths

WWDC26 opens the Foundation Models abstraction. A LanguageModel can back a LanguageModelSession. You care about three conforming types on iOS 27 and macOS 27:

• SystemLanguageModel runs on the device. Lowest latency. Smallest window. No network.
• PrivateCloudComputeLanguageModel runs Apple's larger PCC model. WWDC26 quotes a 32,000 token window and reasoning levels. No API key. Still Apple privacy bounds, not a public LLM vendor.
• CoreAILanguageModel wraps a model you shipped as Core AI (a baked .aimodel). Use it when the job is yours: a style model from the writing LoRA course, or a catalog LLM you chose on purpose.

import FoundationModels
import CoreAI

// Illustrative. Confirm initializers in the installed SDK.
let system = SystemLanguageModel()
let pcc = PrivateCloudComputeLanguageModel()
// CoreAILanguageModel wraps a baked .aimodel. Confirm the load API in Xcode 27.
let coreAI: CoreAILanguageModel? = nil

MLX and third-party server models exist in the same abstraction. This course leaves them out. EDGE FDE routing here is Apple silicon plus Apple PCC plus your Core AI artifact.

Availability is a real branch. The system model can be unavailable. PCC needs a network path that Apple allows. A Core AI file can be missing from the bundle. The router must see those flags before it picks a winner.

Next, write rules in cost, latency, and quality, not in model marketing names.