software-mansion/react-native-executorch v0.5.0

v0.5.0 🚀

on GitHub

Announcing React Native ExecuTorch v0.5.0 ⚛️

What’s new?

🏎️ A full native code rewrite based on C++ JSI bindings enabling zero-copy data transfer between native code and JS
🔥 Up to 2x faster Whisper transcription on iOS
📷 CLIP - extract semantic meaning of images
🚴 Significant improvements to the performance of text embeddings
🧘 Rewrite of the STT streaming API delivering better qualitative results & DX
🐐 Specify LLM structured output schemas using zod
👌 LLMs now correctly handle Unicode/emoji in outputs

⚠️ Breaking changes

• Multiple model instances: JSI bindings remove the static singleton limitation. You can now create and run multiple instances of the same model simultaneously.
• We've replaced separate URL imports with bundled model objects. Instead of manually specifying modelSource, tokenizerSource, and tokenizerConfigSource, just pass a single object like LLAMA_3_2_1B. The individual URLs are still accessible as properties if you need them.

import {
  useLLM,
  LLAMA3_2_1B,
} from 'react-native-executorch';

// Current API:
const llm = useLLM({ model: LLAMA3_2_1B });

// Previous API:
const llama = useLLM({
  modelSource: LLAMA3_2_1B,
  tokenizerSource: LLAMA3_2_TOKENIZER,
  tokenizerConfigSource: LLAMA3_2_TOKENIZER_CONFIG,
});