React Native Local AI (LAI)

A cross-platform React Native mobile application for running Vision Language Models (VLMs) locally on your device.

• iOS: Uses FastVLM (Apple's on-device VLM)
• Android: Uses Gemma (Google's on-device LLM with vision support)

Features

• Cross-platform local AI inference (iOS & Android)
• On-device Vision Language Model execution
• Image+text prompts (image required on iOS, optional on Android)
• Real-time streaming token generation
• Privacy-focused (all processing happens locally)
• Image selection and processing with react-native-image-picker
• Automatic model download and management (Android)
• Dark theme UI with chat interface
• Message persistence with AsyncStorage

Preview

Prerequisites

• Node.js >= 20
• React Native development environment (Setup Guide)
• For Android: Android Studio and JDK 17
• For iOS: Xcode 15+, iOS 18.2+ deployment target, MLX Swift packages

Installation

1. Clone the repository
2. Install dependencies:

yarn install

iOS Setup

1. Install CocoaPods dependencies:

cd ios && pod install && cd ..

2. Download the FastVLM model using the provided script:

cd ios/LAI
./get_pretrained_mlx_model.sh --model 0.5b --dest ./FastVLM
cd ../..

Available model sizes:

• 0.5b - 0.5B parameter model (FP16) - Fastest, smaller size
• 1.5b - 1.5B parameter model (INT8) - Balanced
• 7b - 7B parameter model (INT4) - Most capable, larger size

Android Setup

The Gemma model will be automatically downloaded on first launch.

Running the App

iOS

npx react-native run-ios

Android

npx react-native run-android

Project Structure

LAI/
├── android/              # Android native code (Gemma)
├── ios/                  # iOS native code (FastVLM)
│   └── LAI/
│       ├── FastVLM/      # FastVLM Swift implementation
│       ├── FastVLMModule.swift  # React Native bridge
│       └── FastVLMModule.m     # Obj-C bridge header
├── src/
│   ├── components/       # React components
│   ├── inference/        # VLM inference APIs
│   │   ├── VLMInference.ts      # Cross-platform interface
│   │   ├── FastVLMInference.ts  # iOS implementation
│   │   └── GemmaInference.ts    # Android implementation
│   └── utils/            # Utility functions
└── App.tsx               # Main app component

Usage

1. Launch the app on your device or simulator
2. For iOS: Tap the camera icon to select an image (required)
3. For Android: Optionally select an image with the camera icon
4. Enter your prompt in the text input
5. Tap Send to generate a response
6. Watch the AI response stream in real-time

Clear Chat

Tap the "Clear" button in the header to delete all messages.

Stop Generation

Tap the "Stop" button to cancel an ongoing generation.

Tech Stack

• React Native 0.82
• React 19
• TypeScript
• React Native FS (File system)
• React Native Image Picker
• AsyncStorage
• iOS: FastVLM (MLX Swift, CoreML)
• Android: Gemma (MediaPipe, TensorFlow Lite)

References

iOS

• Apple FastVLM - Apple's Fast Vision Language Model implementation

Android

• MediaPipe LLM Inference for Android - Google's MediaPipe on-device LLM inference guide