Apple Foundation Models in iOS 27: Free On-Device AI for Your App

Apple’s Foundation Models framework picked up multimodal support, dynamic model swapping, and—most concretely—free Private Cloud Compute inference at WWDC 2026. For developers building on iOS 27, that last part changes the cost calculus for adding AI features to apps.

The short version

The Foundation Models framework is Apple’s native Swift API for running on-device and cloud-backed language models. It shipped as a preview in iOS 18 and received a significant push at the 2026 Platforms State of the Union, which Apple described as a “major expansion.”

Three changes matter for production apps:

• The API can now reason about images alongside text.
• You can swap models, tools, and instructions mid-session without losing context.
• Small Business Program members get free Private Cloud Compute inference.

What changed in the Foundation Models framework

The core API gives you a native Swift interface that runs prompts against Apple’s on-device models or, when the task demands it, routes to Private Cloud Compute while keeping data outside Apple’s persistent storage.

With iOS 27, it gains:

Multimodal prompts. Pass images alongside text. Vision framework integration makes OCR and barcode reading available as first-class model inputs—all on-device. A receipt scanner, document parser, or visual search feature now fits within a single Foundation Models session rather than requiring a separate vision pipeline.

Dynamic Profiles. Within a session, you can swap the underlying model, tool set, and system-level instructions without reinitializing. In practice, this means you can run a lightweight model for quick triage and a heavier one for complex synthesis inside the same conversational context—without the user noticing the handoff.

Evaluations Framework. Apple added a dedicated testing layer for AI features. Think XCTest for language model behavior: define expected outputs or behavioral constraints, run the evaluations across varied prompts, and get consistent signal on whether your AI feature holds up under conditions your unit tests don’t cover. This matters because language model outputs are non-deterministic; a test suite that only checks for exact strings will always be fragile.

Free inference: who qualifies

Apple Small Business Program members—developers whose apps have fewer than 2 million first-time downloads in the prior calendar year—now get access to Private Cloud Compute at no cloud API cost. That’s the headline most coverage buried in paragraph five.

For most independent studios and early-stage products, this means zero marginal cost for AI calls that exceed on-device capacity. The pricing floor for AI features in iOS apps just dropped to zero for a large portion of the developer community.

The obvious asterisk: the 2M download threshold means the free tier disappears at the scale where a startup has meaningfully succeeded. At that point, you’re presumably off the Small Business Program and billing on custom enterprise terms. But for the products we’re typically helping get off the ground, the free tier covers everything—and the economics of adding an AI-powered feature look completely different when the incremental cost is nothing.

Multi-provider support: same API, different backends

Apple is opening the Foundation Models framework to third-party model providers via a Language Model protocol in Swift. Claude, Gemini, and any other provider that ships a conforming Swift package can plug in behind the same API surface. You write the call once; the backend becomes a configuration decision.

This matters architecturally for a few reasons:

Fallback chains. Route to on-device Apple models when the device is offline or the task is lightweight. Escalate to a cloud provider when network conditions and task complexity justify it. The routing logic lives in your app, not in a vendor’s SDK.

Model comparisons in staging. Run the same prompts against different backends and compare quality before picking one for production. The uniform API makes this cheap to instrument.

Regulatory constraints. Some markets or enterprise clients specify which cloud infrastructure can process their data. Swapping providers is now a config-level change rather than a re-architecture.

App Intents got sharper edges

Siri AI—rebuilt as a standalone app at WWDC 2026 with support for multi-turn conversation and contextual understanding—is only as capable as what your app surfaces through App Intents. Apple tightened three parts of that API:

Entity Schemas. Contribute typed content to Spotlight’s semantic index. When a user asks Siri to find the invoice they sent last Tuesday, Siri can match against your app’s typed Invoice entities rather than filename strings. The quality of that match is directly proportional to how well you define your schemas.

Intent Schemas. Declare what natural language actions your app supports without enumerating specific trigger phrases. This is a meaningful change from the original App Intents model, where brittle phrase matching was the norm. The system now handles the mapping; you define the semantics.

View Annotations API. Map your views to entities so Siri can reference what’s currently on screen. A user looking at a product page can say “add two of these to my cart” without any disambiguation step—if you’ve annotated the view correctly.

There’s also a new App Intents Testing Framework that exercises Siri, Shortcuts, and Spotlight integration through actual system pathways rather than mocked interfaces. For any app where Siri integration is a feature, this is the first testing path that reflects what the OS actually sees.

What we’re watching

The free inference tier is the story underneath the story. On-device AI has always been compelling in principle—private by default, low latency, no API bill. The limitation was capability. Apple’s cloud tier lifts that ceiling for complex tasks, and pricing it at zero for most of the developer community removes the last objection.

For the products we build, this shifts the conversation. Adding a language model feature used to mean a separate budget line for API costs, a latency estimate, and a privacy disclosure in the App Store listing. On iOS 27, for most client apps, the default answer is on-device first, free cloud fallback, private by design.

The App Intents changes are worth implementing now even if a client isn’t asking for Siri features. Siri AI’s output quality depends entirely on the quality of the metadata your app exposes. Getting that foundational work done now means you benefit automatically from every improvement Apple ships to the underlying model—and given the trajectory since WWDC 2026, those will come often.

If you’re planning a new iOS product or a significant update and want to think through the AI layer from the start, reach out.

Sources

1. What’s New in Apple Intelligence — Apple Developer Documentation
2. Apple Developer News — WWDC 2026 — Apple, June 2026
3. Apple Outlines Major AI and Developer Tool Updates at 2026 Platforms State of the Union — MacRumors, June 9, 2026
4. Apple Siri AI overhaul unveiled at WWDC 2026 with standalone app — The Gadgeteer, June 8, 2026
5. WWDC 2026 Recap — Kodeco, June 13, 2026