Software Mansion's React Native ExecuTorch Runs LLMs On-Device With Zero Cloud Costs
Software Mansion's react-native-executorch lets you drop LLMs, vision, and speech models into any React Native app with a few hooks — no cloud, no native expertise needed.
Takeaways
- What it is: react-native-executorch is an open-source library by Software Mansion for running AI models fully on-device in React Native apps.
- Powered by ExecuTorch: Built on Meta's ExecuTorch runtime, which exports PyTorch models to optimized
.ptebinaries that run locally on iOS and Android. - Broad model support: Covers LLMs (Qwen 3, Llama 3.2, LFM2.5), VLMs, Whisper speech-to-text, Kokoro TTS, OCR, object detection, segmentation, and text embeddings.
- Hook-first API: Everything is exposed as React hooks (
useLLM,useWhisper,useObjectDetection, etc.) with no native or ML expertise required. - Key constraints: Requires React Native New Architecture; no GGUF model support; LLMs need significant device RAM.
- Google Gemma spotlight: The Google Gemma team highlighted the library, noting Gemma 4 on-device support via react-native-executorch is coming soon.
React Native ExecuTorch is an open-source library from Software Mansion that brings full on-device AI inference to React Native apps. It wraps Meta's ExecuTorch runtime , a PyTorch-based framework for deploying models to phones and microcontrollers , behind a clean, declarative React API. The result: LLMs, speech-to-text, computer vision, and more, running entirely on the user's device, with no API calls and no cloud bill.
The library got a wave of fresh attention when the Google Gemma team spotlighted it on social media, noting that Gemma 4 support is on the way. But the library itself has been shipping steadily, and its current feature set is already substantial.
What it actually does
React Native ExecuTorch brings Meta's ExecuTorch AI framework into the React Native ecosystem, enabling developers to run AI models and LLMs locally, directly on mobile devices. It provides a declarative API for on-device inference, allowing you to use local AI models without relying on cloud infrastructure. ExecuTorch achieves state-of-the-art performance through optimizations and delegates such as Core ML and XNNPACK.
The library covers a surprisingly wide surface area of AI tasks out of the box:
- LLMs: chat completion with token streaming, tool calling, and configurable sampling
- Vision-Language Models (VLMs): send images directly to your on-device model for multimodal input
- Speech: Whisper-based speech-to-text with word-level timestamps, plus Kokoro TTS with LLM streaming support
- Computer vision: object detection, instance segmentation, OCR, image classification, and CLIP embeddings