Google's LiteRT.js Brings Native-Speed AI Inference to the Browser
LiteRT.js makes browser-based AI inference practical for real-time tasks, reducing reliance on server-side processing and closing the performance gap with native applications.
Reporting from 1 source: GIGAZINE.
Google released LiteRT.js, a JavaScript runtime that uses WebAssembly to run AI models in browsers at speeds up to 3x faster than previous web runtimes. It supports CPU, GPU, and NPU inference, uses the same .tflite format as LiteRT, and allows partial migration from TensorFlow.js. Real-time demos include object detection, depth estimation, and image upscaling.
LiteRT.js uses the same .tflite model format as Google's existing LiteRT runtime, meaning models already optimized for mobile and desktop can run in a browser without conversion. Developers can also keep pre- and post-processing code in TensorFlow.js while swapping only the inference engine. Google's benchmarks on an M4 MacBook Pro show LiteRT.js up to 3x faster than other web runtimes for vision and audio models, and using WebGPU or WebNN yields 5 to 60x speedups over CPU-only execution. The release includes demos for real-time YOLO object detection, depth estimation from webcam, and 4x image upscaling with Real-ESRGAN, all running client-side.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.