Gemma 4 E2B lands on Pixel 10's Tensor chip
Google's Gemma 4 E2B model now runs natively on the Pixel 10's Tensor chip for offline chat, image and voice tasks, alongside the open Tensor SDK beta.
Google said on July 13 that a lightweight version of its open Gemma 4 model, called Gemma 4 E2B, now runs natively on the custom Tensor chip inside the Pixel 10 line: the Pixel 10, Pixel 10 Pro, Pixel 10 Pro XL and Pixel 10 Pro Fold. The model runs on the phone’s own Tensor Processing Unit rather than a server, so AI chat, image recognition (“Ask Image”), audio transcription (“Ask Audio”), and simple voice or text commands that control phone functions all work fully offline.
Alongside the model, Google pointed developers to the Tensor SDK, which it moved from limited access to open beta in May, paired with the LiteRT toolchain for converting, compiling and deploying models onto the Tensor TPU. The SDK gives developers a model garden of more than 100 classical machine-learning models, a set of precompiled small models on Hugging Face, and the AI Edge Gallery app for testing models directly on Android.
Why it matters
Fully offline AI that never sends prompts, images, or audio to a server is a genuine privacy and reliability upgrade for field or regulated use cases. If you build for Android, the newly open Tensor SDK and free model garden are worth testing before committing to a cloud-only design. It fits the broader push toward on-device models, alongside Ollama’s faster local Gemma 4 and Liquid AI’s phone-sized LFM2.5-230M, though this integration is specific to Pixel’s Tensor chip, not Android generally.