Apple Intelligence at WWDC 2026: Siri AI, On-Device LLMs, and the Post-Cook Era

Apple's 2026 Worldwide Developers Conference, held June 8 at Apple Park under the tagline "All systems glow," was the most AI-heavy keynote in the company's history. For a company that spent two years promising a smarter Siri and delivering mostly delays, this was the moment Apple finally shipped what it had been selling.

Three stories intersect at this WWDC. One is about a product — Siri AI — finally catching up to where users expected it to be three years ago. One is about engineering — how Apple licensed a 1.2-trillion-parameter Google Gemini model and reconciled it with its privacy brand. And one is about transition: Tim Cook's final developer conference as CEO before John Ternus takes over on September 1, 2026.

This article covers what Apple announced, the AI engineering decisions behind it, and what it means for developers and power users going forward.

Siri AI: The Reduilt Assistant That Finally Ships

Apple rebranded Siri to "Siri AI" and acknowledged honestly that "there are times when you expect more from Siri." The rebuilt assistant is the centerpiece of iOS 27 and macOS 27 Golden Gate.

Where it lives: Siri is now embedded in the Dynamic Island (swipe down from it, press the side button, or say "Hey Siri"). It also has a dedicated Siri app on iPhone, iPad, and Mac — conversation history syncs across devices via iCloud, so you can start on your phone and continue on your Mac. On visionOS, Siri gets a 3D visualization you can place anywhere in your space. On the Mac, Siri is built into Spotlight: type a question and get an LLM-generated answer, or right-click any file and have Siri reference it.

What it can do: The demos showed chained, multi-step requests. In one sequence, a presenter asked about a concert, learned booking required a lottery entry, and asked Siri to set a reminder when the lottery opens — and Siri did all three steps. In another, Siri identified a landmark in a photo, pulled up navigation to that location, surfaced photos from a recent family trip, and added a specific image to a shared album on request.

Cross-app synthesis: A presenter asked about a dessert mentioned in a message from a friend. Siri located the relevant details in Messages, compiled them into a watch-party menu, drafted a message to contacts with the menu included, and presented send and edit options. This type of chain-of-action across apps is what makes Siri AI different from the previous version — it can reason across your personal context, not just execute one-shot voice commands.

Siri's voice engine has also been revamped to be more expressive, with micro-adjustable voice settings available during initial setup.

How Siri AI Works: Gemini on the Backend, Foundation Models on Device

The engineering story is the most interesting part of this announcement.

The cloud model: Apple licensed a custom 1.2-trillion-parameter Gemini model from Google, reported at roughly $1 billion per year. This is a mixture-of-experts model: rather than activating all parameters per query, it routes each request to the relevant subset of specialized sub-networks, keeping latency competitive. It's roughly eight times larger than the largest cloud model Apple had previously built in-house.

Where it runs — and the privacy twist: Apple originally planned to run everything on its Private Cloud Compute infrastructure (Apple Silicon servers in stateless, ephemeral sessions). But the 1.2T model was too slow on PCC hardware at Siri's query volumes. So complex queries now route to Google's fleet of Nvidia Blackwell B200 chips on Google Cloud instead — a notable departure from Apple's usual insistence on controlling every layer of the stack. Apple enabled Nvidia's hardware-based confidential computing on those chips, encrypting user input, model weights, and inference results inside GPU memory during computation. Apple's contract with Google also bars Google from using Siri queries to train future models.

The on-device path: Simple requests run on Apple's on-device Foundation Models, which have been updated and are now exposed to third-party developers through the Foundation Models framework — a native Swift API giving apps direct access to the same local LLM that powers Apple Intelligence. This is a privacy-first on-device AI toolkit that lets developers add language understanding and generation to apps without data leaving the device.

Apple also introduced new privacy manifest APIs that let developers declare, on a per-intent basis, whether a Siri interaction may route to the cloud or must stay on-device. For teams handling regulated data, this per-intent declaration is a compliance control, not a UX preference.

Xcode 27: On-Device AI Code Completion

Xcode 27 introduces predictive, multi-line code completion powered by an on-device Apple Intelligence model. It surfaces inline suggestions based on surrounding code context — no cloud round-trip required. This is Apple's most direct answer yet to GitHub Copilot and Cursor, and it runs entirely on Apple Silicon, consistent with Apple's local-first privacy narrative.

Teams that prefer third-party providers can configure Xcode 27 to route suggestions to alternative AI models instead. Beyond completion, the release adds faster simulator performance, tighter Git workflow integration, and improved Instruments profiling focused on memory and energy optimization.

Apple had previously introduced agentic coding in Xcode 26.3 (February 2026) with support for Anthropic's Claude Agent and OpenAI's Codex directly in the IDE. Xcode 27 extends this with Apple's own on-device model for local completions, complementing the cloud-based agentic tools.

App Intents Becomes Mandatory — SiriKit Deprecated

One of the most consequential developer announcements is that App Intents is now the mandatory integration surface for Siri. SiriKit, which has powered voice-assistant features since 2016, received a formal deprecation notice. Apps relying on SiriKit have roughly two to three years before they lose voice-assistant functionality entirely.

The migration is not a rename. SiriKit used XML intent definition files and separate Intent Extensions. App Intents works entirely in Swift: at build time the compiler reads your source and generates compact metadata describing your app's exposed actions, entity types, and query capabilities. The OS can discover what your app can do without launching it. The expanded release adds richer entity types, streaming response support, and structured conversational follow-up.

The strategic implication: apps that do not expose core actions via App Intents become functionally invisible to a Siri that can now chain actions across apps. Apple has already lined up partners including Uber, Amazon, YouTube, WhatsApp, and AllTrails.

iOS 27, macOS Golden Gate, and the OS Lineup

After the divisive Liquid Glass redesign in 2025, WWDC 2026 was a refinement year rather than a reinvention.

iOS 27 supports the same devices as iOS 26 (iPhone 11 and up). Apps launch up to 30% faster. An updated CPU scheduler improves multitasking. Apple rebuilt the on-device search index for faster search across apps, messages, files, and emails.

macOS 27 "Golden Gate" ends support for several older Intel Macs as Apple completes the Apple Silicon transition. Siri is built directly into Spotlight.

Apple Intelligence features now built into the OS:

• Photos: a revamped Clean Up tool, plus Extend (generates extra image around the edges) and Spatial Reframe (uses on-device spatial models to change the angle of a shot)
• Shortcuts: describe the workflow in natural language and Apple Intelligence assembles the shortcut
• Safari: AI tab grouping, page-change monitoring alerts, and vibe-coding browser extensions
• Passwords: automatic password changing
• Notifications: notifications from the same app coalesce into one updating entry
• Visual Intelligence: point your camera or share an image and ask questions about it, available via keyboard shortcut on Mac

Foldable APIs: SwiftUI and UIKit gained new adaptive layout APIs covering hinge-state detection and multi-configuration display handling — quietly laying groundwork for the anticipated iPhone Fold expected this fall.

Tim Cook's Final WWDC and the Ternus Transition

Tim Cook announced in April 2026 that he would become Executive Chairman, with John Ternus (SVP of Hardware Engineering for 25 years) becoming CEO effective September 1, 2026.

Cook's tenure metrics are staggering: market cap from ~$350 billion to $4 trillion, annual revenue from $108 billion to over $416 billion, 2.5 billion active devices, and a services business that alone would be a Fortune 40 company. His final WWDC keynote ended with a brief personal remark acknowledging the transition.

John Ternus has overseen hardware engineering on every product category — iPhone, iPad, Mac, AirPods, Apple Watch. His selection signals continuity: the Apple Silicon transition is complete, the product lines are mature, and the next CEO's challenge is not turnaround but navigation — maintaining Apple's hardware advantage while the industry shifts to AI-driven software.

What This Means for AI Tool Users

For anyone using AI tools professionally, WWDC 2026 changes the competitive landscape in several ways:

Apple enters the assistant race for real. Siri AI, powered by a 1.2T-parameter Gemini model, is now a credible competitor to ChatGPT, Claude, and Gemini on the consumer assistant front. The key differentiator is privacy: Apple's on-device processing means routine AI tasks don't leave your device. For enterprise users handling sensitive data, that matters.

On-device AI becomes a developer tool. Xcode 27's local code completion and the Foundation Models framework for third-party apps mean Apple is betting that the future of AI is hybrid — simple tasks on-device (fast, private, free), complex tasks in the cloud. This mirrors the direction we've seen across the industry: smaller, specialized models running locally, complemented by larger cloud models for heavy lifting.

The Apple-Gemini partnership is a template. Apple using Google's Gemini model, running on Nvidia hardware inside Google Cloud, with confidential computing — this is a remarkable stack for a company that historically owned every layer. It suggests that even Apple cannot afford to build frontier models from scratch and run them at scale. Partnerships in AI infrastructure are becoming as important as chip design or software engineering.

Key Takeaways

• Siri AI is the real product this time: Gemini-powered, cross-app reasoning, conversation history synced across devices, available as a standalone app.
• Xcode 27 ships on-device AI code completion — Apple's direct answer to Copilot and Cursor, with no cloud dependency for basic completions.
• App Intents is now mandatory — SiriKit deprecated with a 2-3 year window; apps without App Intents become invisible to Siri.
• Foundation Models framework opens on-device LLMs to developers — native Swift API, privacy-first, no data off-device.
• Tim Cook's era ends September 1, 2026 — John Ternus takes over as CEO, signaling continuity of Apple's hardware-first strategy.
• The Gemini partnership is a template for Apple's AI future — licensed frontier model, Nvidia confidential compute on Google Cloud, Apple's own on-device models for routine tasks.