Build with the new Apple Foundation Model on Private Cloud Compute

Key Takeaways

• 🔄 Server model is a one-line swap on LanguageModelSession — tools and @Generable stay the same

• 🧠 PCC for reasoning and 32K context; on-device when you need offline with no daily cap

• 💰 Free PCC usage for apps under 2M downloads, no API keys

• 📋 Quota is per-user via iCloud — gate on isAvailable, surface quotaUsage in UI

Introduction

• Private Cloud Compute (PCC) exposes a server LLM through Foundation Models

• Apple’s Server Model and On-device model are accessed through the same APIs

• PCC is for workloads that need more context, reasoning, or heavy tool loops

What is Private Cloud Compute

• Server model; request data isn’t stored or retained after the response

• Wired into OS/iCloud — no auth setup, no API keys, no per-token billing for developers (for apps with under 2M downloads)

• Per-user daily quota (higher tier with iCloud+) for users

• Requires managed entitlement

Integrating PCC with Foundation Models

• Default LanguageModelSession() uses on-devicemodel ; pass PrivateCloudComputeLanguageModel() to switch

• @Generable, tools, and respond signatures are unchanged between models

• PCC features can be gated based on a users available usage quota using model.isAvailable

import FoundationModels

// On-device (default)
let session = LanguageModelSession()
let response = try await session.respond(to: "Summarize this article: \(article)")

// PCC — one-line change
let pccSession = LanguageModelSession(model: PrivateCloudComputeLanguageModel())
let pccResponse = try await pccSession.respond(to: "Summarize this article: \(article)")

let session = LanguageModelSession(
    model: PrivateCloudComputeLanguageModel(),
    tools: [FindRelatedArticlesTool.self]
)

let response = try await session.respond(
    to: "Summarize this article: \(article)",
    generating: ArticleSummary.self  // @Generable works the same
)

On-device vs PCC

Reasoning and context

• Reasoning = extra transcript text before the response; levels: .light, .moderate, .deep

• Set via ContextOptions(reasoningLevel:) on respond

• Reasoning tokens count against contextSize — check the property on each model type

let response = try await session.respond(
    to: prompt,
    contextOptions: ContextOptions(reasoningLevel: .light)
)

SystemLanguageModel().contextSize          // 4096 (26.0), 8192 (27.0)
PrivateCloudComputeLanguageModel().contextSize  // 32768

Usage limits

• Quota is per-user via iCloud, not per-app

• Surface limit state in persistent UI, not alerts

• quotaUsage.status has an approaching-limit case

  • Quota states can be tested through scheme settings to simulate Apple Foundation Model Availabilities

if case .belowLimit(let info) = model.quotaUsage.status, info.isApproachingLimit {
    // warn
}
if model.quotaUsage.isLimitReached {
    // disable + explain
}
if let suggestion = model.quotaUsage.limitIncreaseSuggestion {
    Button("Show options") { suggestion.show() }
}