LiteRT.js: The New JavaScript Runtime Brings Native Hardware Acceleration Via WebGPU, XNNPACK And WebNN, Eliminating Server Costs And Latency Entirely
According to Google Developers Blog, Google launched LiteRT.js on July 9, 2026; a JavaScript runtime that runs AI models directly inside the browser with zero server dependency, bringing its trusted on-device inference library to the web for the first time.
LiteRT.js executes .tflite models entirely client-side, leveraging three hardware backends: XNNPACK for CPU, ML Drift via WebGPU for GPU, and the experimental WebNN API for NPUs. Across real-world computer vision and audio benchmarks, it outperforms existing web runtimes by up to 3x on CPU and delivers 5–60x speedups over standard CPU execution when running on GPU or NPU via WebGPU or WebNN.
The implications go beyond raw speed. Running inference locally means no user data leaves the device, no server bills accumulate, and latency effectively disappears for real-time applications like object detection, audio transcription, and image upscaling. PyTorch models convert in a single step via LiteRT Torch, and Ultralytics YOLO integration ships out of the box.
For web developers, this is the moment edge AI becomes a practical first choice rather than a workaround. The browser just became a serious inference platform.
Check out our previous coverage of AI on The Trusted Times.
