Challenges in Building React Native Applications with On-Device AI
summary
After 1.5 years building on-device features with react-native-executorch the work surfaces concrete constraints: Hugging Face hosts ~3 million models but sizes range from 1 MB to multiple GB so bundling bloats installs and runtime downloads plus a resource fetcher are required for updates. Backends matter: XNNPACK runs on CPU while Core ML/NPU or vendor NPUs give 2x+ speedups, RAM limits (a 350M LFM uses <1 GB) crash mid-tier phones, and continuous inference drains battery unless throttled or prefiltered.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:59Choosing the right model
- 3:46Running models on device
- 5:31Shipping model files
- 8:40Updating models remotely
- 9:24Memory constraints
- 11:07Performance gaps
- 12:13Optimizing inference
- 15:23Fine-tuning models
- 18:18Multi-model pipelines
- 20:23Live demo
related talks
26mBuild your Machine Learning Model on Edge with React Native
10mMachine Learning And React Native
07mMachine Learning on the Edge
17mThink Fast: OnDevice Vector Features
19mHow I Learned React Native with AI
29mA practical approach to building truly native cross-platform applications at scale
24mThe Performance Myths of React Native apps
23m2025: On‑Device AI Goes Mainstream on Android