Google launches LiteRT.js for browser-based AI inference

7/10/2026

LiteRT.js runs AI models in browsers with lower latency and better privacy. Use .tflite models, WebGPU, and WebNN to cut server costs and ship faster.

Google has announced LiteRT.js, a JavaScript binding that brings LiteRT model execution directly into web browsers. The tool is designed to let developers run AI models locally with lower latency, improved privacy, and no serverside inference costs. The release supports existing .tflite models and is positioned as an upgrade path from TensorFlow.js for web deployment. It uses WebAssembly and connects to hardware acceleration paths such as XNNPACK for CPU, ML Drift for GPU through WebGPU, and experimental WebNN support for NPUs. Google says LiteRT.js can be used for tasks such as text generation, object detection, audio processing, depth estimation, and image upscaling. The company also published documentation, an npm package, and demo projects to help developers test the framework and integrate it into browser apps.