What are Apple Foundation Models?

Apple Foundation Models is the family of LLMs that power Apple Intelligence and Siri starting in iOS 27 and macOS 27 Golden Gate. Announced at WWDC 2026 on June 8, the family includes a ~3B-parameter on-device model and a larger server model that runs on Private Cloud Compute. The server model is now reportedly co-trained or fine-tuned with Google's Gemini under the $1B/year licensing deal — Apple uses Gemini's capabilities while keeping inference on Apple infrastructure.

Does Siri send my data to Google in iOS 27?

Officially, no. Apple's design routes Siri requests through Apple Foundation Models running on Apple silicon — either on-device for simple requests or via Private Cloud Compute for harder ones. The Gemini model weights and training are licensed from Google, but inference happens on Apple infrastructure with PCC's no-log guarantees. There's an opt-in 'World Knowledge' fallback to Google's Gemini cloud, similar to the ChatGPT extension in iOS 18+, but that's user-confirmed per session.

Can developers use Apple Foundation Models in their apps?

Yes. iOS 27 ships a new Foundation Models framework that lets developers call Apple's on-device models (and optionally Private Cloud Compute) from their apps with no API key, no metering, and no per-request cost. The on-device model is free to use; PCC requests are subject to fair-use limits. This is Apple's answer to OpenAI's API and Anthropic's Claude SDK — local-first, no data egress, but capped capability.

Quick Answer

Apple Foundation Models + Private Cloud Compute: How It Works

Q: What is Private Cloud Compute?

Private Cloud Compute (PCC) is Apple's verifiable cloud AI service launched in 2024 and expanded for iOS 27. It runs Apple Foundation Models on Apple silicon servers in Apple-controlled data centers. Key privacy properties: requests are not logged, model can't access prior requests, security researchers can independently verify the deployed code, and Apple cannot read individual requests because of stateless cryptographic attestation. PCC is what Apple uses to handle Gemini-grade requests without sending user data to Google.

Published: June 9, 2026

Apple Foundation Models + Private Cloud Compute: How It Actually Works in iOS 27

Announced at WWDC 2026 on June 8: Apple Foundation Models, the LLM family powering iOS 27, macOS 27 Golden Gate, and the new Gemini-augmented Siri. Here’s the full architecture — including how Apple uses Gemini without sending your data to Google.

Last verified: June 9, 2026

TL;DR

Layer	What it does	Where it runs
On-device model (~3B)	Smart Replies, summarization, simple Siri	iPhone Neural Engine
Private Cloud Compute (server)	Complex Siri requests, multi-step reasoning	Apple silicon in Apple data centers
Gemini extension (opt-in)	“World knowledge” queries	Google cloud (opt-in per session)
Foundation Models framework	Developer API for apps	On-device or PCC

The architecture in one diagram (described)

When you ask Siri something in iOS 27:

Phone first asks the on-device model. If the request fits — Smart Reply, summarize, translate, simple action — it never leaves your phone.
If it’s too hard, the request goes to Private Cloud Compute. PCC runs Apple Foundation Models server-side on Apple silicon. No logs. No persistence. Apple itself cannot read the request after it’s processed.
If you want world knowledge (current events, web queries, deep research), Siri asks you per-session if it should escalate to Gemini. That request leaves Apple’s infrastructure to Google.

This is the same pattern Apple introduced with ChatGPT extension in iOS 18.2, generalized.

How Gemini fits in

The $1B/year Gemini deal (covered in Apple Google Gemini Siri $1 billion deal) is for model weights and customized technical support — not for live API calls on every Siri request.

Two patterns:

Distillation/fine-tuning: Apple uses Gemini outputs to train Apple Foundation Models that run on Apple infrastructure. This is the default path. User data never touches Google.
Direct Gemini fallback: Users can opt-in per session to send specific requests to Google’s hosted Gemini for current-event or world-knowledge queries. This is the same model as the ChatGPT extension.

Apple’s framing at WWDC was explicit: “Gemini’s intelligence, Apple’s privacy.”

What is Private Cloud Compute really?

Private Cloud Compute is Apple’s verifiable AI cloud. Key properties:

Apple silicon servers — designed by Apple, no commodity x86 in the trust boundary
Stateless processing — no request is ever written to disk
No retained logs — Apple’s PCC servers can’t be queried later for “what did this user ask”
Cryptographic attestation — every device verifies the PCC node it talks to is running approved code before sending the request
Public code transparency — security researchers can verify the deployed PCC build matches Apple’s published source
Bound to specific request — server can only see one user’s one request, then forgets it

This is the strongest enterprise-AI privacy story currently in market. OpenAI’s enterprise tier and Anthropic’s Claude for Work have similar non-retention promises but none have the cryptographic verification.

What’s actually new in iOS 27

1. Bigger on-device model

The original Apple Foundation Models on-device variant was ~3B parameters. The iOS 27 version is reported to be ~3–4B with significantly better instruction following, thanks to Gemini-derived fine-tuning data.

2. PCC for tool use, not just chat

PCC in iOS 27 supports Siri’s app actions — meaning Siri can call apps, fetch data, take actions inside apps. Previously PCC was mostly summarization and writing tools.

3. Foundation Models framework for developers

This is the big developer story. Apple ships a Swift framework letting any app call:

let response = try await FoundationModels.generate(
    prompt: "Summarize this email in 3 bullets",
    context: emailBody,
    fallback: .privateCloudCompute
)

The on-device call is free. PCC has fair-use limits. Apple takes care of model loading, quantization, and routing.

4. New `WorldKnowledgeProvider` API

Apps can register as a “world knowledge provider” — meaning Wikipedia, news apps, research apps can plug into Siri’s queries. This is the open extension story to the Gemini fallback.

How it compares to competitors

	Apple Foundation Models	OpenAI / Anthropic API	Google Gemini API
On-device default	Yes	No	Limited (Gemini Nano on Pixel)
No data leaves device	Yes (on-device) or PCC (verified)	No	No
Cost per request	$0 on-device, fair-use on PCC	Per-token	Per-token
Capability ceiling	Lower than frontier	Frontier	Frontier
API for developers	New `FoundationModels` framework	Mature	Mature
Available on	iPhone, iPad, Mac	Any platform	Any platform
Cryptographic privacy proofs	Yes (PCC)	Compliance docs only	Compliance docs only

Should developers build on Apple Foundation Models?

Yes, if:

Your app is Apple-platform-first (iPhone, iPad, Mac)
Your users care about privacy
You want to ship AI without billing per request
Your use case fits the on-device capability ceiling (summarization, classification, simple chat)

No, if:

You need frontier reasoning (use Claude Opus 4.8 or GPT-5.5)
Your app needs to run on Android too
You need 1M-token context (use Gemini 3 Pro)
You need real-time multimodal (vision + voice) at frontier quality

What this means strategically

Apple’s play with Foundation Models + PCC + Gemini is the most sophisticated AI strategy among Big Tech in 2026:

Hardware-level privacy moat — Apple silicon + PCC is something no competitor can replicate without spending years on custom silicon
Outsourced frontier intelligence — pay Google $1B/year for capability while keeping the privacy story
Free developer distribution — Foundation Models framework gets billions of users
Optionality — if Google raises prices, Apple’s relationship with Anthropic gives them a hedge (rumored ChatGPT and Anthropic fallbacks also in iOS 27)

Tim Cook’s last big deal sets up John Ternus’s CEO tenure with a 3–5 year AI runway.

Sources

Apple WWDC 2026 keynote (June 8, 2026)
TechCrunch: WWDC 2026 everything announced
The Guardian: Apple debuts Siri AI child safety features WWDC
AppleInsider: Apple Intelligence could bring a better Gemini experience
Marketing AI Institute: Apple $1 billion Siri fix
deeplearning.ai: Apple Gemini partnership analysis