Product Launch · Google Tensor / Pixel

Gemma 4 on the Pixel 10: chat, see, and tweak settings at 30,000 feet, no signal needed

At Google I/O India, the Tensor and Pixel teams showed Gemma 4's lightweight build running directly on the phone's TPU — the chip blocks dedicated to AI — with data staying on the device. Developers can apply for the Tensor SDK to compile and run their own models on Pixel.

Source Google Developers Blog By Prakul Sawhney · Himangshu Roy Where Google I/O India 2026 Devices Pixel 10, 10 Pro, 10 Pro XL, 10 Pro Fold

⚑ Vendor content: Official Google developer blog plus release demo coverage. Videos are live or pre-recorded showcases, not third-party lab tests.

No Wi-Fi at 40,000 feet: does your AI still work?

In July 2026, Google teamed up with the Pixel team at Google I/O India for a joint demo: AI running on the Pixel 10's own silicon. That chip is called Google Tensor, and inside it sits a TPU built specifically for AI workloads.

The pitch: you're on a plane, three miles up, no internet — and you can still have a real conversation. They also walked through the models, showed demo videos, and introduced the tools developers can start using today.

Tensor and Pixel joint demo on stage at Google I/O India
The joint Tensor-Pixel demo at Google I/O India. Image: Google Developers Blog, IMG_5052
Before: cloud-dependent

Ask a question, snap a photo, or record a voice memo, and your phone ships the request to the cloud. No signal, no answer. And even with a connection, your voice and images leave the device.

Now: on-device

As much as possible now runs on the phone's TPU. The model and its compute stay local, and your data never leaves the device.

What Was Announced

Gemma 4
E2B lightweight build, pre-compiled for the Pixel TPU
8 Videos
Official demos: skills, device control, vision, dictation, industry use cases
100+
Classic models on Model Garden, plus a collection of small language models

The official demos keep circling back to four capabilities:

Chat Multi-turn conversations offline, even in an airplane cabin. Runs on the phone's TPU
Vision Point the camera at an object, a plant, or a problem in front of you — no internet needed. Runs on the phone's TPU
Dictation Lectures and voice notes get transcribed on-device, so the data stays put. Runs on the phone's TPU
Settings Toggle Wi-Fi, pull up maps, or trigger other actions with your voice or text. Runs on the phone's TPU

All four share the same TPU, lean toward offline, and keep data on the phone. The fine print lists supported devices: Pixel 10, 10 Pro, 10 Pro XL, 10 Pro Fold.

How Gemma 4 Lite Gets on the TPU

The model is called Gemma 4 E2B for TPU. It's pre-compiled for the Pixel TPU, positioned as lightweight, and designed to run natively on the phone. Pre-built packages are available through the LiteRT community on Hugging Face, with Google_Tensor_G5 in the filename.

If you're curious about Gemma 4 itself — how it was trained, what the parameter count means, and what "thinking" looks like — we've got a detailed breakdown of the technical report:

Related reading
Gemma 4 tech report: How an open-weights model takes on the big players with small parameters, reasoning skills, and memory efficiency
Read that first, then come back for what it can do offline on the Pixel 10.

The 8 Official Demos, Grouped by Capability

Google posted 8 demo videos. Here's how they break down:

1. Offline Skills: Travel, Recipes, Smart Home

The demo shows travel planning, recipe recommendations, and home automation — all without a connection. The product name is Agent Skills.

Agent Skills. Travel / recipe / home assistant demo. Source: Official Google I/O India demo

2. Control Your Phone: Change System Settings by Speaking or Typing

Mobile Actions is powered by Functional Gemma (around 270M parameters, specifically trained for function calling). It runs on the TPU — you command Wi-Fi, maps, and similar features with your voice or text, and the model outputs actions the phone can execute directly.

Mobile Actions · Functional Gemma. Voice / text controls for Wi-Fi, maps, and more. Source: Official Google I/O India demo

3. Chat, See, and Listen — All Offline

Google also split out demos for chat, camera questions, on-device transcription, and a TPU-powered conversation:

ChatOffline Chat
AI Chat: multi-turn conversation offline, including the airplane cabin scenario.
ImageAsk the Camera
Ask Image: identify objects, plants, or on-the-spot problems — no internet required.
AudioOn-Device Transcription
Ask Audio: lectures and voice notes transcribed on the device.
TPUOn-Chip Chat
Edge TPU Chat: conversation inference runs on the Edge TPU.

4. Industry Demos: Retail Store and Auto Garage

The last two videos drop the same offline capability into real-world settings:

RetailRecipe → Store Map
Retail: turn a recipe idea into an in-store shopping map, fully offline.
AutoDiagnose by Photo
Auto repair: snap a photo of a broken part and get a visual diagnosis.

How Developers Get Models Onto a Pixel

The entry point is the Tensor SDK, currently in beta with an application process. The product page lays out three steps:

1 · Compile Compile ahead of time with LiteRT 2 · Deploy Ship via Google Play to Android devices 3 · Infer Run on the TPU inside Google Tensor Source: developers.google.com/edge/tensor-sdk product page
Also worth a look: Model Garden (100+ ready-made models) and small language models pre-built by the LiteRT community.
  1. Apply for the Tensor SDK beta: fill out the form on the Tensor SDK page. Before commercial distribution, go through the licensing and distribution terms.
  2. Grab a ready-made model, or compile your own: find classic models on Model Garden; small language models live in the Google Tensor collection on Hugging Face.
  3. Copy the official samples: check out sample_app_tpu on GitHub, and the Tensor SDK Codelab.
  4. Start with the Gallery app: install LiteRT AI Edge Gallery from the Play Store instead of building a full app first. Supported hardware is the Pixel 10 family.
Source: Unlocking the Next Era of On-Device AI with Google Tensor and Pixel (Google Developers Blog, 2026-07-13, Prakul Sawhney / Himangshu Roy). Product page: Google Tensor SDK.