Google Brings LiteRT GPU Inference to Play Services for Android Apps

Published 29 October 2026 | Publisher: NewsPro | AI-assisted reporting based on the sources below.

Android developers can now access GPU-accelerated LiteRT inference through Google Play services, according to an October 28 announcement from Google’s AI Edge and Android ML teams. The release pairs the hosted runtime with LiteRT’s CompiledModel API, which Google describes as the modern successor to its legacy Interpreter API for GPU acceleration.

A shared runtime, not a bundled copy

Apps that bundle an on-device machine-learning runtime include that code in their own installation package. Google’s new option lets compatible apps obtain LiteRT from Play services instead. Google says this made apps about 8 megabytes smaller in its testing because developers no longer package a separate runtime. That is a company-reported result; actual savings depend on what an app already includes.

The LiteRT documentation describes the toolkit for deploying machine-learning models across devices and platforms. In this release, Google says the CompiledModel API can use CPU or GPU execution, with CPU fallback when a usable GPU is unavailable. A developer currently using the Interpreter API would need to move to CompiledModel to use the new GPU route.

Performance depends on the workload

Google reports GPU inference at roughly 1.4 times the speed of its legacy TensorFlow Lite GPU delegate. The announcement does not present that figure as a result guaranteed across models or devices, so developers would need to measure their own workloads. Model structure and hardware can affect performance.

The update targets developers running custom models; its value depends on compatibility and measured performance.