Questions about local AI on iOS

What is cupertino_fundations_models?

An independent MIT-licensed Flutter iOS plugin that bridges Dart to Apple’s native Foundation Models and Speech frameworks. It exposes local language generation, sessions, streaming, guided structured output and app-defined tools. Start with the English tutorial or guía en español.

Is this an Apple Intelligence Flutter package?

It uses the Foundation Models framework that exposes Apple Intelligence models to applications. It is not an Apple-maintained or Apple-endorsed package and does not enable Apple Intelligence on an unsupported device.

Can I run AI locally on iOS without an API key?

Yes. Select ModelMode.local with CloudPolicy.never. Local generation uses Apple’s system model rather than an external AI API. Your application’s tools may have their own credentials or network behavior.

Does local generation work without internet?

It can run offline after the required Apple Intelligence assets are present and native availability succeeds. Setup and model/Speech asset downloads may require network. PCC, server Speech and networking tools still need a connection.

Do I have to bundle or download an LLM myself?

No custom model file is required by this plugin. Apple manages the system model. The package is not a GGUF, llama.cpp, ONNX or custom-weight runtime.

Which iOS versions are supported?

The plugin deployment target is iOS 15. Foundation Models generation requires iOS 26+ and native Apple Intelligence availability. Some options require iOS 27 or a newer build SDK. Read the compatibility table.

Does it support every iPhone running iOS 26?

No. Hardware eligibility, region, language, enabled settings and model assets also matter. Call checkAvailability() to find out whether the model can run on the user’s device now.

Does the plugin support macOS, Android or web?

No. The declared Flutter plugin platform is iOS. Broader Apple framework availability is not a promise of broader plugin platform support.

Can I import this package into a Swift-only app?

No. This is a Flutter package with a native Swift implementation. For an app written only in Swift, use Apple’s Foundation Models framework directly.

Does this package ship third-party runtime dependencies?

No. Its Dart runtime dependency is Flutter from the SDK; its native bridge uses Apple platform frameworks. Development tooling is separate from runtime dependencies.

Does it support CocoaPods and Swift Package Manager?

Yes. Both plugin integrations use the same Swift sources. Your iOS host still needs an appropriate build SDK, deployment configuration and native rebuild after an upgrade.

Can I request structured JSON rather than prompt for JSON?

Yes. Supply a supported object-root StructuredSchema to generateStructured() or streamStructured(). A prompt that says “return JSON” does not enable native schema guidance. Guided output still needs application validation for facts, amounts, dates and business constraints.

Are streaming events tokens or text chunks to append?

TextSnapshotEvent.text is cumulative. Replace your displayed text with the latest snapshot. The terminal CompletionEvent.response contains the final result. In a structured stream, partial JSON must not be committed as final data.

Can I call my own tools from the model?

Yes. Register tools on SessionOptions. The app validates arguments and checks authorization before side effects. A local model does not make its tools automatically safe or offline. Cancellation does not undo a tool’s side effects.

Is Private Cloud Compute enabled by default?

No. PCC requires its own policy, iOS 27 runtime, Apple-managed entitlement, host signing and provisioning, an opt-in flag, availability and appropriate application consent. Physical PCC verification remains pending. See PCC setup.

Does the plugin automatically fall back to a cloud provider?

No. The consuming application owns external API routing, consent, history, credentials and retries. Failed private prompts are not automatically replayed to another provider. See application-owned routing.

Does on-device generation also make speech-to-text local?

Speech has a separate privacy setting. AudioTranscriptionMode.onDevice is the local default. automatic and server can permit Apple Speech networking. Configure the host usage descriptions and handle denied permissions, missing assets and unsupported locales. Apple Speech servers are not PCC.

Does image input provide native multimodal reasoning?

No. The documented iOS 27 path uses Vision OCR, image classification and barcodes to provide text context. Native Foundation Models image attachments remain disabled because of recorded fatal beta crashes. The package therefore does not offer unrestricted chart, diagram or handwriting understanding.

Is it suitable for long PDFs or financial decisions?

Use bounded input and validate extracted fields against the source. A model response is not proof of complete row coverage or correct amounts. The package does not reconcile financial data, enforce your business rules or persist transactions. See document extraction.

Why does the package name use “fundations”?

The published identifier remains cupertino_fundations_models to preserve existing imports. The human-readable framework name is Apple Foundation Models. Use the exact published identifier in dependencies and Dart imports.