LiteRT.js, Google’s high performance WebAI runtime
summary
LiteRT.js runs converted .tflite models in the browser as an npm package (WASM-backed LiteRT) and targets CPU via XNNPACK and GPU via WebGPU acceleration. Conversion uses AI Edge Torch and AI Edge Quantizer to compress models — one Depth Anything v2 goes from ~100 MB to 27 MB — and a tester shows CPU inference at ~550 ms on CPU while WebGPU can be much faster (GPU failed in one demo). TFJS interop swaps two lines to reuse existing pipelines.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:07What is LiteRT.js
- 2:53LiteRT.js architecture
- 4:18PyTorch to LiteRT
- 6:44Testing and running models
- 8:31Pre-converted models and demos
- 9:55TensorFlow.js interop
- 10:58What’s next for LiteRT.js
- 12:25Closing
related talks
17mRunning Intelligence at the Edge: Web AI in Action
20mTensorFlow.js Bringing Machine Learning to the Web and Beyond
48mTensorFlow.js
02mEasily deploy TF Lite models to the web
41mEscaping the Cloud: High-Performance AI in your Browser
31mDeep Learning at the Edge with TensorFlow.js
46mFrom ML to LLM: on-device AI in the browser
28mMachine Learning on the web with TensorFlow.js