On October 28, 2026, Google announced GPU inference for LiteRT in Google Play services through the CompiledModel API. For Android developers, the key change is that GPU inference requires CompiledModel; an existing Play services Interpreter integration continues to run on the CPU. Google’s announcement also reports smaller apps and faster inference in its testing.

What Google’s performance and app-size figures compare

Google says its tests found GPU inference about 1.4× faster than the legacy TensorFlow Lite GPU delegate. It also reports that apps switching from bundled LiteRT were about 8 MB smaller. Google attributes the size difference to delivering the runtime and GPU accelerator through Play services instead of bundling them in each app. Google’s announcement reports both results as outcomes of its testing.

Interpreter stays on CPU; GPU inference uses CompiledModel

Android Developers’ 2023 walkthrough of TensorFlow Lite in Google Play services and the Interpreter API

An app can keep using the Play services Interpreter API for CPU inference after switching to play-services-litert. To use the announced GPU path, it must use CompiledModel. Google says the legacy GPU delegate and custom delegate APIs are deprecated, and NNAPI removal starts in 2027. Older SDKs receive critical security fixes only.

SituationAPI or requirementResult
Existing Play services Interpreter integrationInterpreterContinues on CPU
GPU inferenceCompiledModelUses the GPU path
Both GPU and CPU requested, with no usable GPURequest both GPU and CPUFalls back to CPU

The 2023 Android Developers walkthrough covers the earlier TensorFlow Lite Interpreter integration, rather than the CompiledModel GPU path.

The migration sequence for an Android app

Google’s LiteRT in Google Play services integration guide and announcement specify this sequence:

  1. Add com.google.android.gms:play-services-litert:17.0.0 or later.
  2. If the app bundles LiteRT, remove its bundled litert dependency.
  3. Initialize LiteRT with GPU enabled and await the resulting environment.
  4. Pass that environment when creating a CompiledModel.

Google’s initialization example is in Kotlin. The model keeps the .tflite format. For GPU processing, LiteRT uses OpenCL where available and OpenGL otherwise. Google says runtime performance improvements, bug fixes and security patches arrive through staged Google Play services rollouts.

Supported platform versions and fallback behavior

Google lists support for Android phones running Android 7 (API level 24) or later, Wear OS 4 or later, and Android Automotive OS on Android 11 or later. If an app requests both GPU and CPU and no usable GPU is available, the runtime falls back to CPU. Google describes NPU support in Play services as in development.