LiteRT in Google Play services gets 1.4x faster GPU inference and smaller apps

Faster GPU inference through the CompiledModel API

Google describes LiteRT in Google Play services as handling more than a trillion inferences a day across billions of devices. The new GPU stack is powered by ML Drift, which Google calls LiteRT's latest open-source GPU engine and says is already used by Google apps on millions of devices daily. It is available through the CompiledModel API, which Google presents as the modern successor to the legacy Interpreter API.

Until now, using this faster GPU stack meant bundling LiteRT inside the app. With the update, the runtime, including the GPU accelerator, ships once with Play services instead of inside every app.

According to Google, the CompiledModel API offers a single workflow for CPU and GPU acceleration instead of a separate delegate and setup path for each accelerator. LiteRT uses OpenCL where available and OpenGL otherwise. If a developer requests both GPU and CPU, it falls back to CPU on devices without a usable GPU. LiteRT keeps the same .tflite model format and runs the same model on Android, iOS, desktop and web.

Smaller apps and wider device support

Google says that in its testing, apps that moved to LiteRT in Google Play services were about 8 MB smaller, which makes them faster to download and lighter on devices. More than 100K Android apps currently bundle their own copy of LiteRT or TensorFlow Lite, so Google expects the savings to add up as more apps switch. Storage and memory savings also grow with each app that shares the runtime.

Google lists further benefits:

• Automatic updates: performance improvements, bug fixes and security patches arrive through Play services in staged rollouts, without an app update.
• More places to run: phones back to Android 7 (API level 24), and now also Wear OS 4+ and Android Automotive OS (Android 11+).
• Tested at scale: ML Kit's Vision and Text APIs, such as Subject Segmentation and Document Scanner, already run on LiteRT in Google Play services.

The Scrapbook feature in the Google AI Edge Gallery app, where image segmentation runs on LiteRT in Google Play services, serves as a demonstration.

How to switch, and what is deprecated

Google says the move takes a few lines of code. Developers add com.google.android.gms:play-services-litert:17.0.0 (or a later version), remove the litert dependency if they bundle it today, get the LiteRT environment from Play services and pass it in when creating a CompiledModel. The rest of the code stays the same.

Apps already using the legacy Interpreter API in Google Play services keep working on CPU after switching to play-services-litert. Running on the GPU requires moving to the CompiledModel API.

Google also introduced a LiteRT skill for Android Studio. In Agent Mode, developers can ask it to "migrate my TFLite code to CompiledModel" or type @litert-runtime. It switches the app to play-services-litert, rewrites Interpreter calls and sets up GPU acceleration with CPU fallback. Google advises treating the result as a first pass: review the diff, build, and compare outputs on a real device.

The legacy GPU delegate and custom delegate APIs are deprecated in favor of CompiledModel GPU acceleration and the Compiler Plugin and Dispatch APIs. NNAPI will be removed starting in 2027, and older SDKs now receive critical security fixes only. Shipped apps keep working, but Google warns that code using deprecated APIs will not compile against the latest SDK.

What it means for Android developers

For developers running custom models on CPU or GPU, Google encourages switching to LiteRT in Google Play services, and the deprecations make planning a move to the CompiledModel API necessary. Google says LiteRT will keep receiving regular updates, including better GPU acceleration and wider model coverage, and that NPU support in Play services is actively being worked on.

Developers who run into problems can reach the team through GitHub Discussions or by filing an issue.