Artur Morys-Magiera

This talk digs into what it really takes to run LLMs efficiently on mobile hardware in a React Native environment. We’ll examine the constraints teams face - memory limits, model loading strategies, inference performance, platform-specific APIs - and how they shape real-world product decisions. From there, we’ll introduce a React Native library that provides two complementary ways to integrate on-device AI: cross-platform, state-of-the-art models that run locally on both Android and iOS, and a dedicated path for leveraging Apple Intelligence capabilities on supported iOS devices.
We’ll walk through the architecture, usage patterns, and trade-offs of each approach, and discuss best practices for delivering smooth, low-latency AI experiences without relying on the cloud. We will cover both native libraries for mobile AI & LLM model inference, and the available wrappers for React Native, along with the trade-offs, capabilities, hardware compatibility, model format compatibility and compile-time model optimizations (operator fusing, vectorization, memory planning, operations accelerated for specific hardware) considerations so that the audience is aware to pick the best solution for their specific use cases.

Many mobile teams reach a point where maintaining duplicate implementations across iOS and Android becomes unsustainable. React Native promises a unified approach - but for mature apps, a full rewrite is rarely an option. It’s expensive, risky, and difficult to align with continuous product development. In practice, for mature products, most migration efforts are discarded as infeasible. React Native Brownfield is the solution - a set of tools for embedding your React Native app in native apps as a self-contained library, which allows for a gradual adoption of React Native in native applications.In this talk, we will do a deep dive into the internals of Brownfield: how are RN apps packaged, what tooling is needed to cover all use cases (navigation stack, JS-native state sharing, CLI), why we had to implement a custom Gradle Plugin, a few things about real-life problems with transitive dependency handling, SO stripping, how Hermes V1 is handled on Android, how prebuilt RN core on iOS (0.81+) is handled with backwards compatibility, and finally - how the integration of this complex library with Expo is handled using just 1 line of code for the end user, thanks to an Expo config plugin.