On July 9, 2026, Google introduced LiteRT.js, a new member of the LiteRT family designed to execute machine-learning models directly in the web browser. Its main acceleration path uses WebGPU, with WebNN planned as the platform evolves and WebAssembly available as a CPU fallback.
For web engineering this is a meaningful shift. Some AI workloads can run without continuously sending data to a remote server: local classification, signal processing, selected computer-vision tasks or interface personalisation are possible candidates. That can reduce latency, lower backend cost and, in the right use case, keep more user data on the device.
Browser AI still has real constraints. Teams must account for model size, download time, memory, thermals, battery use and large capability differences between desktop and mobile browsers. A production feature also needs a graceful fallback instead of assuming every device exposes the same acceleration APIs.
For product teams, runtimes such as LiteRT.js expand the architecture toolbox. The question is no longer only which model to use, but where inference should happen: cloud, private infrastructure, the user’s device, or a hybrid of all three.