What are Microsoft's Aion models?

Aion is a family of small in-house models announced at Microsoft Build 2026 (June 2, 2026) as part of the seven-model MAI release. Aion runs locally on Windows PCs — NPU-accelerated on Copilot+ PCs and CPU-fallback on standard Windows 11. It is Microsoft's answer to Apple Foundation Models and on-device LLM workflows. Aion includes a text model and Aion Vision for multimodal.

How does Aion compare to Apple Foundation Models?

Both are small on-device LLM families designed for OS-level integration and developer SDKs. Apple Foundation Models ship in iOS 27, iPadOS 27, and macOS 27 with the Apple Intelligence SDK; they run on Apple Silicon NPUs. Aion ships on Copilot+ PCs (Snapdragon X, AMD Ryzen AI 300, Intel Core Ultra 200V) with the Windows AI Foundry SDK. Capability is comparable for autocomplete, summarization, classification. Apple has 12 months more polish; Microsoft has broader hardware reach.

Which should I build on as a developer?

If you ship a cross-platform desktop app, build on both — they expose similar primitives. If you ship a Mac/iOS app, Apple Foundation Models has the better developer story and tighter OS integration. If you ship a Windows-first or enterprise app, Aion gives you direct access on Copilot+ PCs with Azure AI Foundry continuity to the cloud MAI family. Neither replaces frontier models — for hard tasks, you still route to Claude Fable 5 or GPT-5.5.

Do these models work offline?

Yes, that is the point. Both Aion (on Windows) and Apple Foundation Models (on Apple Silicon) run entirely locally. No internet required after install. This is critical for privacy-sensitive workloads, regulated industries, and consumer apps that want to advertise on-device-only AI. The on-device tier handles autocomplete, basic summarization, classification, simple Q&A. Frontier reasoning still needs cloud.

Quick Answer

Aion on Windows vs Apple Foundation Models: On-Device AI (June 2026)

Published: June 12, 2026

Aion on Windows vs Apple Foundation Models: On-Device AI

Two of the biggest OS announcements of 2026 are about on-device AI. Apple shipped Foundation Models with iOS 27 / macOS 27 in June after WWDC 2026. Microsoft announced the Aion small model family at Build 2026 on June 2. Here is the on-device AI comparison for developers in mid-2026.

Last verified: June 12, 2026

TL;DR

Capability	Microsoft Aion (Windows)	Apple Foundation Models
Announced	June 2, 2026 (Build 2026)	June 2026 (WWDC)
Runs on	Copilot+ PCs (Snapdragon X, AMD Ryzen AI 300, Intel Core Ultra 200V) + Windows 11 CPU fallback	Apple Silicon (M1+, A17+, A18+, A19)
Developer SDK	Windows AI Foundry	Apple Intelligence SDK / Foundation Models framework
Multimodal	Aion Vision	Apple Vision (Foundation Models multimodal)
Offline	Yes	Yes
Cloud continuity	Azure AI Foundry → MAI-Thinking-1	Apple Private Cloud Compute
Best for	Windows-first apps, enterprise	Mac/iOS apps, consumer privacy

What Microsoft shipped on June 2

The MAI launch at Build 2026 included Aion as the on-device family:

Aion text — short-form reasoning, summarization, autocomplete.
Aion Vision — multimodal text + image on-device.
Targets — Copilot+ PCs (Qualcomm Snapdragon X Elite/Plus, AMD Ryzen AI 300, Intel Core Ultra 200V with NPU). CPU fallback on standard Windows 11.
SDK — Windows AI Foundry, integrated into Visual Studio and VS Code.
Cloud continuity — when on-device confidence is low, the SDK can route to MAI-Thinking-1 in Azure AI Foundry. Developers control routing policy.

See Microsoft MAI-Thinking-1 vs GPT-5.5 vs Claude Fable 5 for the full MAI family context.

What Apple shipped at WWDC 2026

Apple Foundation Models were the WWDC 2026 headline. Key facts:

Ships in iOS 27, iPadOS 27, macOS 27, visionOS 27.
Available through the Foundation Models framework for third-party developers.
Runs entirely on Apple Silicon NPU.
Cloud continuity via Apple Private Cloud Compute — Apple-controlled servers with verifiable privacy guarantees.
Native integration with App Intents, Shortcuts, system-wide Writing Tools, and Genmoji.

For the broader WWDC 2026 picture, see WWDC 2026 vs Google I/O 2026 vs Microsoft Build 2026 keynote recap and Apple Foundation Models vs Anthropic Agent SDK vs OpenAI Agents SDK.

Capability comparison

Both families are small (single-digit billion parameter range) and designed for OS-level tasks. Public benchmarks at this tier are sparse, but the realistic capability set:

Task	Aion	Apple Foundation Models
Autocomplete (text)	✅	✅
Summarization (1–10 page)	✅	✅
Tone rewriting	✅	✅ (Writing Tools)
Classification	✅	✅
Short Q&A	✅	✅
Image description	✅ (Aion Vision)	✅
Hard reasoning	❌ — route to cloud	❌ — route to cloud
100K+ token context	❌ — route to cloud	❌ — route to cloud
Agentic multi-step	Limited	Via App Intents only
Coding tasks	Via GitHub Copilot integration	Limited (Xcode 27 Foundation Models integration)

For Xcode-specific developer flows, see Xcode 27 Foundation Models Claude vs Gemini vs OpenAI.

Privacy model

Element	Aion	Apple Foundation Models
On-device default	Yes	Yes
Cloud fallback	Azure AI Foundry (MAI family)	Apple Private Cloud Compute
User opt-in for cloud	Developer-controlled	System-controlled with explicit consent
Auditability	Microsoft audit logs	Cryptographic verification of PCC node
Data residency	EU/US selectable	Apple-controlled global

Apple’s Private Cloud Compute story is more developed — see Apple Foundation Models vs Private Cloud Compute. Microsoft’s Azure compliance is more mature for enterprise data residency.

Hardware reach

Platform	Compatible hardware (June 2026)
Aion	Copilot+ PCs (~5%+ of Windows installed base, growing fast), Windows 11 CPU fallback (broadest reach)
Apple Foundation Models	iPhone 15 Pro/Pro Max and newer, M1+ Macs, A17+ iPads, Apple Vision Pro

Apple has higher per-device performance ceiling. Microsoft has broader installed-base reach via CPU fallback.

Developer story

Building on Aion

Install Windows AI Foundry SDK.
Pick a model variant (Aion text, Aion Vision).
Define routing policy: on-device only, cloud fallback, or always-cloud.
Ship through Microsoft Store or any Windows distribution channel.

Building on Apple Foundation Models

Adopt the Foundation Models framework in Xcode 27.
Define App Intents for system-level access.
Privacy review by App Store for cloud-routed features.
Ship through App Store.

Apple has tighter OS integration (Writing Tools, Genmoji, system-wide Shortcuts). Microsoft has more flexible enterprise deployment.

Who should build on what

Build on Aion if:

Your app is Windows-first or cross-platform.
You need enterprise deployment with Azure data residency.
You already use GitHub Copilot and want continuity.
You target Copilot+ PCs as the premium tier.

Build on Apple Foundation Models if:

Your app is Mac/iOS-first.
You need the system-level App Intents integration.
Your users care about visible privacy guarantees (PCC).
You target the iPhone Pro / Apple Silicon Mac premium tier.

Build on both if:

You ship a cross-platform desktop or note-taking app.
Both expose similar primitives — autocomplete, summarization, tone rewrite, classification, vision.

What you still need cloud for

Neither family handles:

1M-token context (use Claude Fable 5 or GPT-5.5).
SWE-Bench Pro-level coding (use Claude Fable 5).
Multi-hour agent loops (use Claude Fable 5 or Claude Code).
Multimodal video understanding at scale (use Gemini 3.5 Pro).

On-device is the floor; frontier cloud is the ceiling.

Bottom line

Apple Foundation Models for tight OS integration on Mac/iOS. Microsoft Aion for Windows-first apps with Azure continuity. Both run offline, both have cloud fallback, both target the same on-device task tier. Cross-platform apps should build on both. Frontier workloads still go to Claude Fable 5, GPT-5.5, or Gemini 3.5 Pro in the cloud.

Sources: Microsoft Build 2026 keynote (June 2, 2026), Apple WWDC 2026, blogs.microsoft.com, developer.apple.com.