Google Upgrades LiteRT for Faster Android Inference

Google AI Edge Team and Android ML Team announced on October 28, 2026, that LiteRT in Google Play services now supports GPU inference about 1.4x faster than the legacy GPU delegate. The Android on-device inference runtime handles more than a trillion inferences a day across billions of devices. The upgrade uses ML Drift through the CompiledModel API, which Google describes as the modern successor to the legacy Interpreter API.

Google says apps switching from a bundled runtime to LiteRT in Play services were about 8 MB smaller in its testing. The runtime and GPU accelerator ship once with Play services instead of inside each app; more than 100K Android apps currently bundle LiteRT or TensorFlow Lite. Google says shared runtime use can save storage and memory as more apps switch, while performance improvements, bug fixes and security patches arrive through staged Play services rollouts without an app update. LiteRT supports phones running Android 7 (API level 24) and later, Wear OS 4+ and Android Automotive OS on Android 11+.

LiteRT retains the .tflite model format and offers one workflow for CPU and GPU acceleration. It uses OpenCL where available and OpenGL otherwise, and falls back to CPU if a requested GPU is unavailable. The same model format is supported on Android, iOS, desktop and web. The legacy Interpreter API remains supported on CPU, allowing phased migration. Google cites ML Kit Vision and Text APIs, including Subject Segmentation and Document Scanner, as services already using this infrastructure across billions of Android devices.

To adopt it, developers add com.google.android.gms:play-services-litert:17.0.0 or later, remove the bundled litert dependency if present, initialize the Play services LiteRT environment, and pass it to CompiledModel. Existing Interpreter code continues to work on CPU after switching; GPU use requires moving to CompiledModel. Google also offers an Android Studio skill that can rewrite Interpreter calls and configure GPU acceleration with CPU fallback, while advising developers to review its changes, build, and compare outputs on a real device.

Google says the legacy GPU delegate and custom delegate APIs are deprecated in favor of CompiledModel GPU acceleration and the Compiler Plugin and Dispatch APIs. NNAPI removal begins in 2027; older SDKs receive critical security fixes only, and code using deprecated APIs will not compile against the latest SDK. Google says shipped apps continue to work, but recommends planning migration. NPU support in Play services is being worked on.

Google AI Edge Team and Android ML Team announced on October 28, 2026, that LiteRT in Google Play services now supports GPU inference about 1.4x faster than the legacy GPU delegate. The Android on-device inference runtime handles more than a trillion inferences a day across billions of devices. The upgrade uses ML Drift through the CompiledModel API, which Google describes as the modern successor to the legacy Interpreter API.

Google says apps switching from a bundled runtime to LiteRT in Play services were about 8 MB smaller in its testing. The runtime and GPU accelerator ship once with Play services instead of inside each app; more than 100K Android apps currently bundle LiteRT or TensorFlow Lite. Google says shared runtime use can save storage and memory as more apps switch, while performance improvements, bug fixes and security patches arrive through staged Play services rollouts without an app update. LiteRT supports phones running Android 7 (API level 24) and later, Wear OS 4+ and Android Automotive OS on Android 11+.

LiteRT retains the .tflite model format and offers one workflow for CPU and GPU acceleration. It uses OpenCL where available and OpenGL otherwise, and falls back to CPU if a requested GPU is unavailable. The same model format is supported on Android, iOS, desktop and web. The legacy Interpreter API remains supported on CPU, allowing phased migration. Google cites ML Kit Vision and Text APIs, including Subject Segmentation and Document Scanner, as services already using this infrastructure across billions of Android devices.

To adopt it, developers add com.google.android.gms:play-services-litert:17.0.0 or later, remove the bundled litert dependency if present, initialize the Play services LiteRT environment, and pass it to CompiledModel. Existing Interpreter code continues to work on CPU after switching; GPU use requires moving to CompiledModel. Google also offers an Android Studio skill that can rewrite Interpreter calls and configure GPU acceleration with CPU fallback, while advising developers to review its changes, build, and compare outputs on a real device.

Google says the legacy GPU delegate and custom delegate APIs are deprecated in favor of CompiledModel GPU acceleration and the Compiler Plugin and Dispatch APIs. NNAPI removal begins in 2027; older SDKs receive critical security fixes only, and code using deprecated APIs will not compile against the latest SDK. Google says shipped apps continue to work, but recommends planning migration. NPU support in Play services is being worked on.