Repose

An offline, private AI chat assistant for React Native. It runs a small language model (Llama 3.2 1B) entirely on-device using ExecuTorch — no network, no cloud, your data never leaves the phone.

Features

• Streaming chat — tokens appear live as the model generates them
• Fully offline — works in airplane mode; no API keys or servers
• Model load progress — live percentage while the .pte model loads
• Live metrics — tokens/sec, token count, and latency per response
• Tiny footprint — int4 quantized 1B model (~700 MB)

Tech

Layer	Choice
App	React Native CLI, TypeScript
Runtime	react-native-executorch
Model	Llama 3.2 1B Instruct, int4 quantized to .pte
Export	Python + torch.export + ExecuTorch (scripts/export_llama.py)
Native bridge	NativeModule + NativeEventEmitter for token streaming

Getting started

npm install
npm run export-model        # produces models/llama3_2_1b_q4.pte
npm start                   # metro
npm run ios                 # or npm run android

Bundling the .pte as a native asset and the native module setup follow the react-native-executorch docs.

How it works

1. src/modules/ExecuTorchLLM.ts wraps the native module and surfaces onToken / onLoadProgress / onError events.
2. src/hooks/useLLM.ts manages model lifecycle and accumulates streamed tokens into the response.
3. src/screens/ChatScreen.tsx renders bubbles, the progress bar, and the metrics strip.