Apple's first on-device foundation model gave developers a new execution path for AI features. iOS 27 expands what the system can do, while Android is moving in the same direction. If you last looked at this space around the initial release, the trade-offs and possible products have already changed.
System-provided models have a distinctive place in the stack. They ship with the device, are optimized for its hardware, work offline, keep data local, and remove the need to package and maintain a model yourself. Their capabilities are defined by the platform and differ from cloud inference and models supplied by the application.
This talk reviews the additions announced at WWDC and our work to bring the same approach to React Native on Android. Our integrations are built around the Vercel AI SDK, so applications can use familiar APIs regardless of where inference runs. We will look at what system models can now do, where their constraints appear, and which product features are particularly well suited to them.
If your last experiment stopped at the initial Apple Intelligence release, this talk will give you a reason and a path to try again.
This talk has been presented at React Summit US 2026, check out the latest edition of this React Conference.
