Apple Siri AI & AFM 3: The Five-Model On-Device Privacy Architecture That Changes Everything

Executive Summary

Apple's WWDC 2026 (June 8) marked the most significant transformation of Siri since its 2011 debut. Siri AI is not an incremental update — it is an entirely new assistant built on the third-generation Apple Foundation Models (AFM 3), a family of five custom-built models that span from a 3-billion-parameter on-device dense model to a 20-billion-parameter sparse on-device MoE, to three server-based models including one running on NVIDIA GPUs in Google Cloud via an extended Private Cloud Compute (PCC) architecture.

The architectural innovation is the story. Apple's AFM 3 Core Advanced — a 20B-parameter model that runs on-device — uses a novel Instruction-Following Pruning (IFP) technique to store the full model in NAND flash and selectively load only 1–4B parameters into DRAM per prompt, breaking the traditional DRAM-wall that limits on-device model size. This is fundamentally different from standard MoE routing (which swaps experts token-by-token) — Apple makes routing decisions per prompt, using a lightweight dense block to select a fixed set of experts during initial processing and periodically reselecting them during generation.

The AFM 3 Cloud Pro model — Apple's most capable server-based model for agentic tool use and complex reasoning — runs on NVIDIA GPUs in Google Cloud, marking the first time Apple has extended PCC to third-party infrastructure. The security architecture includes a cryptographically verifiable, append-only ledger of all Google Cloud hardware, dual independent roots of trust, and layered protections that treat every component from firmware through application code as part of the trusted computing base.

Siri AI brings personal context understanding (accessing messages, emails, photos), onscreen awareness, broad world knowledge via web search, expressive voices, and higher-accuracy dictation. A dedicated Siri app syncs conversation history across all Apple devices via iCloud. Visual Intelligence expands from iPhone to iPad, Mac, and Apple Vision Pro.

Key finding: Apple has built the most comprehensive on-device AI stack in the industry — not by chasing parameter counts, but by rethinking the hardware-software boundary (NAND-based expert routing), extending privacy guarantees to third-party infrastructure (PCC on Google Cloud), and integrating AI so deeply into the OS that it becomes invisible. The trade-off is clear: no open weights, no API access for third-party developers beyond the Foundation Models framework, and a walled-garden distribution model that only reaches Apple devices.

---

1. The AFM 3 Family: Five Models, One Privacy Architecture

Apple's approach to the AFM 3 family is fundamentally different from the single-flagship strategy of OpenAI, Anthropic, or Google. Instead of one massive model doing everything, Apple shipped five specialized models optimized for different hardware and use cases:

Model	Parameters	Architecture	Hardware	Primary Use
AFM 3 Core	3B (dense)	Dense transformer	All Apple Intelligence devices	General text, image understanding
AFM 3 Core Advanced	20B total, 1-4B active	Sparse (IFP-based)	iPhone 17 Pro/Air, M3+ Mac, M4+ iPad, Vision Pro	Expressive voices, dictation, multimodal
AFM 3 Cloud	Not disclosed	PT-MoE (upgraded)	Apple silicon servers	Server-side reasoning, multimodal
ADM 3 Cloud (Image)	Not disclosed	Diffusion	Apple silicon servers	Image generation, editing, Genmoji
AFM 3 Cloud Pro	Not disclosed	Not disclosed	NVIDIA GPUs (Google Cloud)	Agentic tool use, complex reasoning

1.1 The Common Foundation

All five models "shared a common initial foundation before specializing for their respective architectures and use cases." This means they were trained from the same base checkpoint and then specialized — AFM 3 Core for on-device efficiency, AFM 3 Core Advanced for sparse on-device inference, AFM 3 Cloud for server-side performance, ADM 3 Cloud for image generation, and AFM 3 Cloud Pro for agentic reasoning.

The training data mix includes "publicly available information, data licensed or purchased from third parties, open-sourced data, data obtained through dedicated studies, and synthetic data." Apple explicitly states: "We do not use our users' private personal data or user interactions when training our foundation models."

---

2. AFM 3 Core Advanced: Breaking the DRAM Wall

2.1 The Problem

Traditional large language models — whether dense or sparsely activated — require all weights to reside in active memory (DRAM). This creates a massive footprint that limits scalability on consumer hardware. A 20B dense model requires approximately 40GB of VRAM/DRAM (at FP16), which exceeds the unified memory of most iPhones and many iPads.

Standard MoE models solve this by activating only a subset of experts per token, but they still require the full model to be loaded into DRAM for routing — the router needs to see all expert weights to make decisions.

2.2 Apple's Solution: Instruction-Following Pruning

Apple's approach, based on their research paper "Instruction-Following Pruning for Large Language Models", fundamentally changes the memory model:

Key architectural differences:

Dimension	Standard MoE	AFM 3 Core Advanced (IFP)
Model storage	All weights in DRAM	Full model in NAND flash
Routing granularity	Per token	Per prompt (with periodic reselection)
Expert selection	Router sees all weights	Lightweight dense block selects fixed set
Memory footprint	~40GB DRAM for 20B	~4GB DRAM (1-4B active + shared)
Data movement	Token-by-token expert swap	Prompt-level expert swap
Shared experts	None (or minimal)	High percentage of always-active shared experts

2.3 Inference-Time Elasticity

The architecture introduces "inference-time elasticity" — rather than using a single model for all tasks, AFM 3 Core Advanced uses a predetermined number of active parameters tailored to each specific use case. Weights are loaded incrementally across requests of varying difficulty, scaling the model size far beyond traditional DRAM limits while minimizing latency.

This means a simple query (e.g., "What's the weather?") might activate only 1B parameters, while a complex reasoning task might activate 4B — all from the same 20B model stored in flash.

2.4 Device Requirements

AFM 3 Core Advanced is "unlocked by and optimized for our most capable Apple silicon systems":

• iPhone Air, iPhone 17 Pro, iPhone 17 Pro Max
• iPad (M4) or later with at least 12GB unified memory
• Mac (M3) or later with at least 12GB unified memory
• Apple Vision Pro (M5)

Older Apple Intelligence devices (iPhone 15 Pro, iPhone 16 base models, M1/M2 Macs) get AFM 3 Core (3B dense) instead.

---

3. The Server-Side Stack: AFM 3 Cloud, ADM 3 Cloud, and the Google Cloud Extension

3.1 AFM 3 Cloud: The Workhorse

AFM 3 Cloud represents a substantial step forward in multimodal reasoning, built on an upgraded version of Apple's Parallel-Track Mixture-of-Experts (PT-MoE) architecture introduced in 2025. Key improvements include:

• Stabilized training for larger context windows
• Improved reasoning over long contexts with accurate recall
• Multimodal capabilities including image understanding and audio
• Optimized for Apple silicon servers in Apple data centers

3.2 ADM 3 Cloud (Image): The Diffusion Model

The image model powers:

• Image Playground — high-quality image generation in virtually any style, including photorealistic
• Advanced photo editing — Spatial Reframing, Extend, and touch-based modifications
• Genmoji — custom emoji generation
• Specialized adapters for downstream editing experiences

The model "generalizes across different aspect ratios and resolutions" and draws on the broader AFM family to guide both creation and editing.

3.3 AFM 3 Cloud Pro: The Google Cloud Extension

This is the most controversial and technically interesting part of the architecture. AFM 3 Cloud Pro — Apple's most capable server-based model — runs on NVIDIA GPUs in Google Cloud, marking the first time Apple has extended PCC to third-party infrastructure.

3.3.1 The Security Architecture

Apple and Google collaborated to build capabilities "that go far beyond a traditional confidential computing deployment":

Key security guarantees:

1. No reliance on confidential computing alone — Apple treats every component from firmware through application code as part of the trusted computing base
2. Cryptographically verifiable hardware ledger — an append-only record of all Google Cloud hardware in the PCC fleet
3. Dual roots of trust — critical components are attested against at least two separate roots of trust from independent vendors
4. Short-lived inference software — shared inference software is recycled with a short time-to-live duration
5. Isolated key management — attested keys are held in a separate, dedicated confidential VM isolated from external inputs

3.3.2 Why This Matters

This extension is significant for three reasons:

1. It breaks Apple's "all Apple silicon" rule — Previously, PCC ran exclusively on Apple silicon servers. Extending to NVIDIA GPUs in Google Cloud is a pragmatic acknowledgment that some workloads (agentic tool use, complex reasoning) require more compute than Apple's own data centers can provide at scale.

2. It establishes a template for third-party PCC — The security architecture (hardware ledger, dual roots of trust, full-stack attestation) could become a standard for any company wanting to run privacy-preserving AI on third-party infrastructure.

3. It creates a dependency on Google — Apple is now dependent on Google Cloud infrastructure for its most capable AI model. This is a reversal of the traditional power dynamic, where Google would typically be the competitor.

---

4. Siri AI: The Experience

4.1 What's New

Siri AI is "an entirely new version of Siri" with capabilities that go far beyond the previous voice assistant:

Capability	Previous Siri	Siri AI
Conversation	Single-turn, command-based	Multi-turn, conversational, context-aware
Personal context	Limited (calendar, contacts)	Messages, emails, photos, notes, files
World knowledge	Pre-trained cutoff	Live web search for up-to-date information
Onscreen awareness	None	Can see and reason about what's on screen
Visual Intelligence	iPhone Camera only	iPhone, iPad, Mac, Apple Vision Pro
Dedicated app	No	Yes — with iCloud-synced conversation history
Writing tools	No	Compose, rewrite, proofread, summarize (system-wide)
Expressive voices	No	Yes (AFM 3 Core Advanced)
Dictation	Basic	Higher accuracy (AFM 3 Core Advanced)

4.2 The Dedicated Siri App

For the first time, Siri gets a dedicated app where users can:

• Revisit past conversations
• Start new conversations
• Sync conversation history across all Apple devices via iCloud
• Continue a thread started on Mac on iPhone, iPad, Apple Watch, or Apple Vision Pro

This is a significant shift from the previous model where Siri was only accessible via voice activation or the Control Center.

4.3 Visual Intelligence Expansion

Visual Intelligence — the ability to point your camera at something and get information — expands from iPhone to:

• iPad — onscreen content analysis
• Mac — onscreen content analysis
• Apple Vision Pro — objects in the world (spatial awareness)

4.4 System-Wide Integration

Siri AI is integrated into:

• Spotlight — Answer questions directly from search
• Messages — Contextual suggestions and Smart Reply matching user's writing style
• Mail — Contextual suggestions that trigger actions in third-party apps
• Photos — Spatial Reframing, Extend, Image Playground
• Safari — Tab grouping by topic, webpage monitoring, custom extensions from natural language
• Home — Batch camera notifications, search footage by description
• Maps — Enhanced Flyover with AI-generated aerial imagery
• Shortcuts — Natural language automation assembly

---

5. The Developer Story: Foundation Models Framework

5.1 The Framework

Apple's Foundation Models framework is a native Swift API that gives developers direct access to Apple Foundation Models — on device and in Private Cloud Compute — as well as any model provider conforming to the Language Model protocol.

Key capabilities:

• Multimodal prompts — Reason about images alongside text
• Dynamic Profiles — Swap models, tools, and instructions within a continuous session
• Evaluations framework — Ensure AI features work reliably across dynamic conditions
• On-device Vision framework — Image reasoning alongside text
• No cost per request — On-device models are free to use

5.2 App Intents

The App Intents framework connects apps to Apple Intelligence and Siri AI through schemas:

• Entity schemas — Contribute app content to the Spotlight semantic index for personal context understanding
• Intent schemas — Enable natural language actions on app content without specific phrases
• View Annotations API — Map views to entities for on-screen awareness

5.3 The Walled Garden

Unlike OpenAI's API (accessible to anyone with a credit card) or Anthropic's API (broad third-party integration), Apple's Foundation Models framework is:

• iOS/macOS/visionOS only — No web API, no cross-platform access
• Swift-only — No Python, no REST API
• Apple-controlled — No self-hosting, no fine-tuning access
• Free for on-device — But server-side models have daily usage limits (increased with iCloud+)

This is a deliberate trade-off: deeper integration and better privacy at the cost of developer freedom.

---

6. Performance Evaluations

Apple conducted extensive human evaluations of the AFM 3 models. Key results:

6.1 General Text Capabilities

Model	Preferred vs. 2025 Baseline	Improvement
AFM 3 Core	45.6% (vs. 23.3% for 2025 baseline)	96% relative improvement
AFM 3 Cloud	64.7% (vs. 8.7% for 2025 baseline)	644% relative improvement
AFM 3 Cloud Pro	~10% relative improvement over AFM 3 Cloud	Incremental but meaningful

6.2 Image Understanding

Model	Preferred vs. 2025 Baseline
AFM 3 Core	61%+ (when users preferred one over the other)
AFM 3 Cloud	37.8% (vs. 9.6% for 2025 baseline)
AFM 3 Cloud Pro	~14% relative improvement over AFM 3 Cloud

6.3 Dictation

AFM 3 Core Advanced demonstrates a positive win rate in overall quality compared to Apple's existing production dictation system, with preference extending consistently across all individual formatting and comprehension dimensions.

6.4 Overall Response Satisfaction

• AFM 3 Cloud: ~36% relative improvement in overall response satisfaction over 2025 baseline
• AFM 3 Cloud: ~21% relative improvement in instruction following over 2025 baseline
• AFM 3 Cloud Pro: ~10% relative improvement in text, ~14% in image understanding over AFM 3 Cloud

Important caveat: These are vendor-reported human evaluations, not independent benchmarks. Apple does not report scores on SWE-Bench, GPQA, MMLU, or any standardized benchmark that would enable direct comparison with GPT-5.6, Opus 4.7, or Gemini 3.5 Pro.

---

7. The Privacy Architecture: Private Cloud Compute

7.1 The Three Guarantees

Apple's PCC architecture is built on three guarantees:

1. Your data is never stored — Requests are processed and discarded; no logs, no training data collection
2. Used only for your requests — Compute resources are dedicated to your request and recycled
3. Verifiable privacy promise — Third-party security researchers can independently verify the security architecture

7.2 How It Works

7.3 The EU and China Restrictions

• EU: Mac and Apple Vision Pro users can access Siri AI. iOS, iPadOS, and watchOS users cannot initially due to Digital Markets Act constraints. Apple states it is "working hard to find a path forward that preserves its users' privacy and security."
• China: Siri AI and Apple Intelligence features will not be available while Apple works through regulatory requirements.

---

8. Positioning in the 2026 Frontier

8.1 The Complete Picture

8.2 Apple's Unique Position

Apple occupies a unique position in the 2026 landscape:

Dimension	Apple	OpenAI	Anthropic	Google	Microsoft
On-device model	Yes (3B-20B)	No	No	Limited (Gemini Nano)	No
Privacy guarantees	PCC (verifiable)	Standard	Standard	Standard	Standard
Open weights	No	No	No	Partial (Gemma)	No
Third-party API	No (walled garden)	Yes	Yes	Yes	Yes
OS integration	Deep (system-wide)	None	None	Android (limited)	Windows (Copilot)
Hardware control	Full (silicon to OS)	None	None	Partial (TPU)	Partial (Maia 200)
Distribution	2B+ devices (estimated)	API + ChatGPT	API + Claude	Google Search + Gemini	Azure + Copilot

8.3 When to Use Apple's Stack

Apple's stack is the default when:

• You're in the Apple ecosystem (iPhone, Mac, iPad, Vision Pro)
• Privacy is the primary concern (on-device processing, verifiable PCC)
• You need deep OS integration (Spotlight, Messages, Photos, Shortcuts)
• You want AI that works offline (on-device models)

Look elsewhere when:

• You need open weights → Qwen3.6-27B, DeepSeek V4 Pro, Kimi K2.7 Code
• You need a third-party API → OpenAI, Anthropic, Google, Microsoft
• You need the strongest reasoning → GPT-5.6, Opus 4.7, Fable 5
• You need cross-platform deployment → Any cloud API provider
• You need fine-tuning → OpenAI, Anthropic, or open-weight models

---

9. Comparison with Recent Articles

Our prior analysis established the frontier landscape that Apple now enters:

Dimension	MAI-Thinking-1 (Microsoft Mai Model Family Frontier Tuning Hill Climbing Machine 2026 06 17)	Qwen3.7 Max (Qwen37 Max Plus Closed Weight Frontier Agent Era 2026 06 16)	Gemini 3.5 Flash (Gemini 35 Ecosystem Flash Pro Audio Antigravity 2026 06 15)	Apple AFM 3
On-device model	No	No	Limited (Nano)	Yes (3B-20B)
Silicon co-design	Maia 200	No	TPU	Apple silicon
Privacy architecture	Standard	Standard	Standard	PCC (verifiable)
OS integration	Deep (Windows)	None	Android (limited)	Deep (iOS/macOS)
Open weights	No	No	Partial	No
Custom tuning	Yes (Frontier Tuning)	No	No	No
Third-party API	Yes	Yes	Yes	No (walled garden)
Benchmark transparency	Vendor-reported	Independent (AA Index)	Vendor-reported	Vendor-reported only

Apple's unique differentiators are the on-device capability (20B sparse model on iPhone), the verifiable privacy architecture (PCC with third-party audit), and the deep OS integration (system-wide AI). The weaknesses are the lack of benchmark transparency (no independent scores), the walled-garden distribution (Apple devices only), and the no third-party API (developers limited to Foundation Models framework).

---

10. Key Takeaways

1. Apple has built the most comprehensive on-device AI stack in the industry. The AFM 3 Core Advanced (20B sparse, 1-4B active) running on iPhone is something no other company has achieved at this scale. The Instruction-Following Pruning technique — storing the full model in NAND and selectively loading experts into DRAM — is a genuine architectural innovation.

2. The five-model family is a deliberate strategy. Rather than one model doing everything, Apple optimizes each model for its specific hardware and use case: 3B dense for broad device coverage, 20B sparse for flagship devices, server-side workhorse for complex tasks, diffusion model for images, and the Cloud Pro for agentic reasoning.

3. The Google Cloud PCC extension is pragmatic and risky. Running AFM 3 Cloud Pro on NVIDIA GPUs in Google Cloud is a necessary compromise for performance, but it creates a dependency on Google and requires a sophisticated security architecture to maintain privacy guarantees. The dual roots of trust and cryptographically verifiable hardware ledger are impressive, but they represent a significant expansion of Apple's trusted computing base.

4. Siri AI is a ground-up rebuild, not an update. The dedicated app, iCloud-synced conversation history, personal context understanding, onscreen awareness, and system-wide integration represent a fundamental rethinking of what a digital assistant can be. The expansion of Visual Intelligence to iPad, Mac, and Vision Pro completes the cross-platform story.

5. The walled garden is both strength and weakness. Deep OS integration and verifiable privacy are advantages that no cloud API can match. But the lack of a third-party API, Swift-only developer access, and Apple-device-only distribution limit the ecosystem's reach and innovation potential.

6. Benchmark transparency is a gap. Apple reports only vendor-conducted human evaluations. Without scores on SWE-Bench, GPQA, MMLU, or the AA Intelligence Index, it's impossible to compare AFM 3 Cloud Pro against GPT-5.6, Opus 4.7, or Gemini 3.5 Pro on standardized metrics.

7. The EU and China restrictions create a fragmented experience. The Digital Markets Act standoff and Chinese regulatory requirements mean Siri AI will not be available to hundreds of millions of potential users — at least initially.

---

11. References & Resources

• Apple Newsroom: WWDC 2026 Keynote
• Apple Newsroom: Siri AI Announcement
• Apple Machine Learning Research: Third-Generation AFM
• Apple Security: Expanding Private Cloud Compute
• Apple Developer: Foundation Models Framework
• Apple Intelligence Product Page
• Apple Research: Instruction-Following Pruning
• Apple Research: PT-MoE Architecture (2025)
• Microsoft Mai Model Family Frontier Tuning Hill Climbing Machine 2026 06 17 — Microsoft MAI full-stack comparison
• Qwen37 Max Plus Closed Weight Frontier Agent Era 2026 06 16 — Qwen3.7 Max closed-weight strategy comparison
• Gemini 35 Ecosystem Flash Pro Audio Antigravity 2026 06 15 — Gemini 3.5 ecosystem and TPU co-design comparison
• Ai News Week 2026 06 08 2026 06 15 — Weekly roundup covering WWDC 2026 announcements

---

12. Future Directions

Several questions remain open:

• AFM 3 Cloud Pro benchmarks: When will Apple release scores on standardized benchmarks (SWE-Bench, GPQA, MMLU, AA Intelligence Index)? Without these, the "most capable server-based model" claim remains unverified.
• EU Siri AI timeline: How long will the Digital Markets Act standoff last? Will Apple find a technical solution, or will EU iOS/iPadOS/watchOS users remain excluded indefinitely?
• China regulatory path: What specific regulatory requirements is Apple working through? Will Siri AI launch in China this year, or is it a multi-year effort?
• Open-weight AFM: Will Apple ever release open weights for any AFM model? The 3B Core model would be a compelling community contribution, but Apple's walled-garden strategy suggests this is unlikely.
• Foundation Models framework adoption: How many third-party apps will integrate Apple Intelligence via the Foundation Models framework? The Swift-only, iOS-only constraints limit adoption compared to OpenAI's broad API.
• AFM 4 timeline: With Apple's established yearly WWDC cadence, when will the fourth generation arrive, and what capabilities will it bring?
• PCC on other clouds: Will Apple extend PCC to AWS, Azure, or other cloud providers beyond Google Cloud? The security architecture could become a standard for privacy-preserving cloud AI.
• Siri AI vs. Google Assistant: With Gemini 3.5 Pro approaching and Google's own on-device models improving, how will the Siri AI vs. Google Assistant competition evolve?
• Apple's AI revenue model: How will Apple monetize Apple Intelligence? Will it be bundled with iCloud+ subscriptions, or will there be a separate AI subscription tier?
• The Google dependency: Apple's reliance on Google Cloud for AFM 3 Cloud Pro creates an unusual dependency. Will Apple build its own data center capacity for these workloads, or will the Google partnership deepen?

---

Article published: June 18, 2026, 11:25 AM SGT Status: Draft — pending build and commit