In brief

  • Size and GPU figures are results reported by Google.
  • Developers must integrate the runtime.
  • Announced NPU support is still to come.

A shared runtime, not a new app

On October 28, Google announced improvements to LiteRT in Google Play Services. This is a runtime that can execute machine-learning models inside an application, not a new chatbot for users to download.

The approach shares that runtime across applications rather than bundling a full copy into each one. Google reports a reduction of up to approximately 8 MB per app in its tests. That does not mean every application on your phone will automatically become smaller.

The numbers need context

The company also reports GPU performance up to 1.4 times that of its previous GPU delegate. This is the supplier’s comparison, not a promise covering every model, phone or task.

Developers need to integrate the solution and check their results. Listed compatibility starts with Android 7, Wear OS 4 and Android Automotive 11. Devices without Google Play Services need another route, such as the standalone runtime.

On-device does not mean no communication ever

The documentation says inputs used by these APIs, such as images or text, are processed on the device and are not sent to Google for that operation. It also describes the collection of performance metrics and communication with servers for updates.

It would therefore be misleading to describe every app using it as completely private or always offline. NPU acceleration in Play Services is still a future announcement. For now, the practical news is a technical option that can improve applications once their developers incorporate it.

Sources and context

Sources used for this article. Follow the links to read the original material.

Topics in this article

AndroidGoogleArtificial intelligence

See the linked sources for more information. Corrections and contact →