Google is pushing a more concrete on-device AI workflow into developers' hands. In a new LiteRT post, the company highlighted production apps already using NPU acceleration and pointed to a related update in Google AI Edge Gallery, its reference app for testing local models on phones.

What changed

Google's April 24 release notes for AI Edge Gallery 1.0.12 say the app now adds an NPU option for Gemma 3 1B inside Prompt Lab on supported Qualcomm Snapdragon devices. For Play Store installs, Google says the needed libraries are bundled automatically on compatible phones. The release also lists supported Snapdragon 8 Gen 2 through 8 Elite device families from Samsung, OnePlus, Xiaomi, ASUS, Motorola, Honor and others. Google separately says AI Edge Gallery includes built-in benchmarking tools, giving developers a straightforward way to compare local model performance across hardware paths.

The accompanying developer blog gives the broader context. Google says LiteRT is already used in products including Google Meet, Epic's Live Link Face beta for Android, and Argmax's Android speech stack, all aimed at keeping inference fast without overheating the device or draining battery.

Why it matters

The conservative takeaway is not that every phone can suddenly run frontier models well. Google's own release limits the new NPU path to specific Qualcomm hardware and a small Gemma model. But it does make on-device testing more tangible: developers can now try CPU, GPU, and NPU paths in Google's own app instead of stitching together a custom demo first.