Rybczak describes the package as a way to bring Apple Foundation Models to React Native. Because the models ship with iOS, they can be reused across apps and are designed for iPhone memory needs. He contrasts their typical 300 to 400 MB app memory use with up to 4.5 GB of RAM for a third-party model on iPhone 15 Pro.
Intelligent Fallbacks: Using Apple Intelligence Seamlessly With @react-native-ai/apple
Speaker
Szymon Rybczak, a senior React Native developer at Callstack, explains how to use Apple Foundation Models in React Native through the @react-native-ai/apple package on React Native AI Unpacked.
Setup involves installing the library and CocoaPods, then rebuilding the native app. To test in a macOS simulator, Rybczak says developers need macOS Tahoe and iOS 26 simulators. He points to the package documentation for setup details and simulator troubleshooting.
The demo uses the Vercel AI SDK with the Apple provider’s model, including its streamText API. With networking disabled, the app answers a question about React Native on the device. Rybczak describes the response as smooth and fast, with no internet connection needed.
Rybczak also demonstrates structured output using a Zod schema with name and description string fields. The model returns an object matching those fields, which the app can then display or use in its interface.
Rybczak recommends trying the package in an iOS 26 app and consulting the documentation for API guidance and troubleshooting. He directs developers to the Callstack Incubator AI repository and invites them to report issues on GitHub.
Takeaways
- Apple’s built-in models can be shared across apps, which may reduce the memory burden compared with third-party models.
- Test simulator support against the documented macOS Tahoe and iOS 26 requirements.
- The Vercel AI SDK can stream responses from Apple’s on-device model while the network is disabled.
- A Zod schema can shape model output into fields an app can use directly.
Mentioned in this talk
Search inside this talk
What was claimed 12
- role atSzymon Rybczak is Senior React Native developer at
Callstack“I'm a senior React Native developer at Callstack.”
YouTube Verified6 sources· Talk
- featureApple LLM: Shipped with iOS itself
“it's shipped with iOS itself”
YouTube Unverified1 source· Talk
- integrates withReact Native AI Apple integrates with React Native
“brings Apple Foundation Models to React Native”
YouTube Unverified1 source· Talk
- integrates withReact Native AI Apple integrates with Apple LLM
“brings Apple Foundation Models to React Native”
YouTube Unverified1 source· Talk
- built
“React Native AI Apple is our pretty new library”
YouTube Unverified1 source· Talk
- metriciPhone 15 Pro: Maximum third-party model RAM per app is 4.5 GB
“which is 4.5 GB for iPhone 15 Pro”
YouTube Unverified1 source· Talk
- metricReact Native AI Apple: Built-in models usually consume 300 to 400 MB on the app itself
“usually 300 to 400 MB on the app itself”
YouTube Unverified1 source· Talk
- uses
“backed by the Apple LLM”
YouTube Unverified1 source· Talk
- featureReact Native AI Apple: Supports text streaming and structured outputs through the Vercel AI SDK
“streamText, generateObject, and other good stuff from the Vercel AI SDK”
YouTube Unverified1 source· Talk
- featureReact Native AI Apple: Generates responses on device without an internet connection
“the response will be generated on the device”
YouTube Unverified1 source· Talk
- featureReact Native AI Apple: Supports schema-constrained structured output with name and description string fields
“I will force the AI, the LLM, to match the schema”
YouTube Unverified1 source· Talk
- usesReact Native AI Apple uses Zod
“I define it with a Zod object”
YouTube Unverified1 source· Talk
