LiquidChat - a showcase of on-device AI capabilities using Apple Foundation Models, MLC LLM Engine and Vercel AI SDK.

Demo.

• On-device LLMs via Apple Foundation Models or MLC LLM Engine.
• AI-generated chat topics and group conversations.
• Light and Dark themes with flexible customization.
• CI/CD with GitHub Actions and EAS.
• Modularized architecture with Dependency Injection.

• For Apple Foundation Models, MacOS 26 with Xcode 26 are required.
• No above restrictions for MLC models. However, testing is limited.

Install Bun.

bun i

cp .env.example .env
# (Optionally) update the EXPO_PUBLIC_AI_FALLBACK_MODEL

Start Metro bundler and follow the instructions in terminal to run the app.

bun run start

Pre-built MLC models aren't available for Android making the platform not testable at the moment. Follow callstackincubator/ai for updates.

MLC models require a physical device, making them not testable on Simulator. As a workaround, it's possible to run on Mac or a physical device.

Mac / Physical device:

• Build Xcode project: npx expo prebuild > open ios/liquidchat.xcworkspace
• Adjust Signing & Capabilities with your Personal Team

Simulator:

• Remove MLC deps: bun rm @react-native-ai/mlc > npx expo prebuild --clean > then adjust the ai/index to always return AppleAIProvider.

Schiffradar

Artur Yersh