Apple's On-Device AI vs Real Local LLMs: What WWDC 2026 Actually Changed

WWDC 2026 opened on 8 June with Tim Cook's final keynote as CEO. The headline was a rebuilt AI strategy: a new dedicated Siri app with iCloud-synced conversation history, six OS releases in beta (iOS 27, iPadOS 27, macOS 27, watchOS 27, tvOS 27, visionOS 27 — full launch targeted fall 2026), and a homeOS developer preview for a forthcoming HomePad smart-home hub.

The AI layer is called Apple Intelligence, now co-developed with Google using Gemini technology. The on-device models (AFM Core / AFM 3 Core Advanced) are Apple's own. The cloud model (AFM 3 Cloud Pro) is refined using Gemini outputs and runs on Nvidia GPUs in Google Cloud.

Apple Intelligence routes each task through one of three tiers depending on complexity. Which tier a task hits determines the privacy story.

The most misunderstood part of WWDC 2026 is the Google relationship. Apple distinguishes 'trained using Gemini' from 'is Gemini'. The on-device models — AFM Core and AFM 3 Core Advanced — are Apple's own and have no Google involvement. Your on-device interactions never go to Google.

The cloud model (AFM 3 Cloud Pro) is different. It runs on Nvidia GPUs in Google Cloud. Apple states the model was refined using Gemini outputs — a knowledge-distillation step where Gemini's outputs served as training signal. The result is Apple's own model, but hosted on Google infrastructure.

Reported (unconfirmed): the partnership is worth approximately $1B/year; the cloud model is reportedly around 1.2T parameters. Apple reportedly attempted its own Private Cloud Compute hardware for heavy tasks first but found it too slow, leading to the Google Cloud arrangement.

Apple's on-device model and a self-hosted open-weight LLM both process on local hardware — but the differences are significant:

The practical question: does my data stay on my device? The answer depends entirely on which tier handles the task. Apple provides some transparency, but you cannot directly observe which tier fires for any given request.

The developer story from WWDC 2026 is less about model quality and more about surface area. Apple is extending App Intents so Apple Intelligence can call into third-party apps — but only through explicitly declared actions and data structures. Siri won't scrape your UI; it calls declared intents.

This is functionally parallel to GEO (Generative Engine Optimization). Instead of structuring content for AI search crawlers, you're structuring the action surface your app exposes to the OS model. Apps that declare clean, granular App Intents will appear in Apple Intelligence results; apps that don't, won't.

For EU/GDPR-regulated businesses: the on-device tier provides data residency by default, which may satisfy Article 32 obligations for simple tasks. The Cloud Pro tier routes to Google Cloud in the United States, raising the same Chapter V cross-border transfer questions that apply to any US cloud service. Legal teams should assess whether Apple Intelligence falls within DPIA scope for their workload.

Apple just made 'private, on-device AI' a mainstream expectation for roughly a billion device owners — that validation of the local-first premise matters. But Apple Intelligence is a hybrid, partly Google-backed, closed-weight system: a gateway to the local-AI mindset, not a replacement for running your own models.

If privacy is your primary motivation, the three-tier architecture introduces real caveats: cloud tiers process your requests, the Cloud Pro tier runs on US-based Google Cloud infrastructure, and you control none of the weights, routing logic, or update schedule.

Self-hosted local LLMs — Qwen, Llama, Gemma on your own hardware — remain the only architecture where you can verify that nothing leaves your environment. Apple's on-device tier narrows the gap significantly for everyday tasks, but the 'trust Apple' requirement never disappears.

For EU users: on-device gives you data residency on simple tasks. For complex tasks routed to Google Cloud, the same GDPR Chapter V analysis applies as for any other US cloud service.