Features & Use cases
OCR
Extract text from any image, receipt, or document – right on the device. No image ever leaves the phone, which makes it a natural fit for sensitive documents: contracts, medical records, financial statements.
With React Native ExecuTorch, you can ship a fully functional document scanner in your app without routing user data through an external OCR API.
Computer vision
Run object detection, segmentation, and image classification. React Native ExecuTorch supports models including YOLO, RF-DETR, MobileNet, and Segment Anything (SAM), giving you a full computer vision toolkit.
From background blur in video calls to product recognition in retail apps – real-time object detection is the feature end users will notice.
Voice capabilities
Add studio-quality speech synthesis and accurate transcription to your app. React Native ExecuTorch ships with Whisper for speech-to-text and Kokoro for text-to-speech, so users can listen to content or dictate input even when they're offline.
Build an audiobook reader, a voice note app, or an accessibility tool that works even on a plane.
Image & text embedding
Generate vector embeddings from images and text – enabling semantic search, similarity matching, and local RAG without a cloud vector database.
A great fit for note-taking apps, photo managers, or knowledge tools – with the entire embedding and retrieval pipeline kept local.
LLMs & VLMs
Run large language models and vision-language models. React Native ExecuTorch supports Llama, Phi, and other popular open models.
Build AI-powered chat, smart autocomplete, or image-aware assistants – for apps where privacy isn't a feature checkbox, but the entire product.
OCR + LLM Pipeline
Chain OCR and an LLM together on-device: snap a photo of a sign, menu, or document in any language – React Native ExecuTorch extracts the text and feeds it to a local language model.
For travel apps, document assistants, or translation tools – like Google Lens, but private and offline.
Developers are already shipping with it
Expo
@expo
🔥 The future of AI apps is on the device. With react-native-executorch (from @swmansion), you can run models locally in your Expo app: ♢ Private by design ♢ Zero API fees ♢ Works offline ♢ Instant response Read how it works in the blog below from @_darthez_ and @Nklockiewicz ↓
Bran Aust
@bran_aust
Built a real-time voice conversation loop in @expo. Speech-to-text runs locally using @swmansion's react-native-executorch + Whisper. Transcribed text is sent to a local model, generates a response, then gets converted to audio via OpenAI TTS and played back with expo-audio. All on-device. Fast + seamless.
Resolver Vicky | Dev
@resolvervicky
Someone built a way to run AI models directly on your phone using React Native. No server. No API calls. No internet required. React Native ExecuTorch by Software Mansion.
Expo
@expo
🔥 The future of AI apps is on the device. With react-native-executorch (from @swmansion), you can run models locally in your Expo app: ♢ Private by design ♢ Zero API fees ♢ Works offline ♢ Instant response Read how it works in the blog below from @_darthez_ and @Nklockiewicz ↓
Bran Aust
@bran_aust
Built a real-time voice conversation loop in @expo. Speech-to-text runs locally using @swmansion's react-native-executorch + Whisper. Transcribed text is sent to a local model, generates a response, then gets converted to audio via OpenAI TTS and played back with expo-audio. All on-device. Fast + seamless.
Resolver Vicky | Dev
@resolvervicky
Someone built a way to run AI models directly on your phone using React Native. No server. No API calls. No internet required. React Native ExecuTorch by Software Mansion.
Roboflow
@roboflow
seeing more and more vision models being deployed on mobile repos like React Native ExecuTorch are making it easier for agents to one-shot vision apps, much needed now
ExecuTorch team
@Executorch
Success stories: Ecosystem Integration Hugging Face: Optimum-ExecuTorch for direct transformer model deployment LiquidAI: Next-generation Liquid Foundation Models optimized for edge deployment Software Mansion: React Native ExecuTorch bringing edge AI to mobile apps
Thor 雷神 ⚡️
@thorwebdev
#Gemma 4 on-device in React Native 💎🔥 Huge shoutout to @swmansion for making this happen! Y'all are awesome 🫶
Roboflow
@roboflow
seeing more and more vision models being deployed on mobile repos like React Native ExecuTorch are making it easier for agents to one-shot vision apps, much needed now
ExecuTorch team
@Executorch
Success stories: Ecosystem Integration Hugging Face: Optimum-ExecuTorch for direct transformer model deployment LiquidAI: Next-generation Liquid Foundation Models optimized for edge deployment Software Mansion: React Native ExecuTorch bringing edge AI to mobile apps
Thor 雷神 ⚡️
@thorwebdev
#Gemma 4 on-device in React Native 💎🔥 Huge shoutout to @swmansion for making this happen! Y'all are awesome 🫶
Google Gemma
@googlegemma
Gemma 🤝 React Native 📱 Exciting news for mobile developers! We love seeing the community unlock new ways to build. You'll soon be able to run Gemma 4 completely on-device in React Native.
Vegetable_Sun_9225
@Vegetable_Sun_9225
It does, it performs very well, probably best performing framework for mobile/edge right now. Comparing directly with a PC is kinda rough since it's hardware dependent. A PC with an RTX 4090 is going to blow away the theoretical performance max on an iPhone, the wattage alone is at least an order of magnitude. I'm getting 10 t/s for llama 3.1 8b on an S24+ You can easily get your hands wet by following the instructions on torchchat which gives you demo apps for Android and iOS or checkout the react native executorch wrapper. Both are pretty easy to get rolling with.
Google Gemma
@googlegemma
Gemma 🤝 React Native 📱 Exciting news for mobile developers! We love seeing the community unlock new ways to build. You'll soon be able to run Gemma 4 completely on-device in React Native.
Vegetable_Sun_9225
@Vegetable_Sun_9225
It does, it performs very well, probably best performing framework for mobile/edge right now. Comparing directly with a PC is kinda rough since it's hardware dependent. A PC with an RTX 4090 is going to blow away the theoretical performance max on an iPhone, the wattage alone is at least an order of magnitude. I'm getting 10 t/s for llama 3.1 8b on an S24+ You can easily get your hands wet by following the instructions on torchchat which gives you demo apps for Android and iOS or checkout the react native executorch wrapper. Both are pretty easy to get rolling with.
See what's possible
An AI personal assistant built entirely with React Native ExecuTorch
Private Mind is a production-grade AI assistant app that runs entirely on your phone. It uses React Native ExecuTorch under the hood for LLM inference, speech recognition, and on-device embeddings.
What's new from the RN ExecuTorch team?
Check out our latest blogpost, conference talks and release notes
On-device AI beats cloud for TTS - here's why
Running Kokoro TTS on-device cuts per-request cloud costs, drops latency, and keeps audio private.
AI-powered note-taking in React Native - Part 1: Text semantic search
Semantic search over notes with on-device text embeddings - meaning, not keyword matches.
AI-powered note-taking in React Native - Part 2: Image semantic search
Extend search to images with multimodal embeddings that share space with text queries.
AI-powered note-taking in React Native - Part 3: Local RAG
Add a local Retrieval-Augmented Generation pipeline so users chat with their notes fully offline.
AI-powered note-taking in React Native - Part 4: Speech recognition
On-device speech-to-text closes the loop - talk to the assistant, keep audio on the device.
Top 6 local AI models for privacy and offline use
Six on-device models worth shipping when privacy, cost, and offline capability matter.
FAQ
On-device AI runs inference entirely on the user's phone. It means no server round-trips, no API costs, and no user data ever leaving the device. Your app works offline, responds faster (no network latency), and keeps all sensitive data private by default.
React Native ExecuTorch supports a range of open-source models across multiple categories: YOLO and RF-DETR for object detection, Segment Anything (SAM) for image segmentation, Whisper for speech-to-text, Kokoro for text-to-speech, Llama and Phi for large language model inference, and CLIP-style models for image and text embeddings. See the full list of available models on our Hugging Face.
Yes – all inference happens locally on the device. Once the model is bundled with or downloaded to the app, it runs completely offline. No internet connection is needed for any inference task, including LLM chat, speech recognition, image processing, or embedding generation.
Edge AI is a broader category that covers any inference running outside a central cloud server – including on-device phones, IoT devices, and edge servers. On-device AI specifically refers to inference running on the end user's device (smartphone, tablet, laptop). React Native ExecuTorch is built for iOS and Android – the model runs on the phone itself, not on a nearby server.
Yes. React Native ExecuTorch is open source and maintained by Software Mansion. It's available on GitHub under an open-source license. Commercial use is permitted – see the repository for the full license terms.
We built the library so most teams can ship on-device features themselves with the APIs it provides. For teams with tighter timelines, more complex pipelines, or custom model requirements, we also work directly with companies to design, train, and integrate on-device AI into production React Native apps. If that sounds useful, get in touch.