Product launch · Google Tensor / Pixel

Google runs Gemma 4 on Pixel 10: chat, vision, and settings changes at 30,000 feet offline

At Google I/O India, the Tensor and Pixel teams showed a light Gemma 4 build running on the phone’s TPU (the chip block made for AI math), with data staying on the device. Developers can apply for the Tensor SDK to compile models for Pixel.

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

⚑ Vendor content: Google’s developer blog plus a launch-demo recap. Clips are showcase footage, not third-party lab tests.

Plane is offline—can AI still work?

In July 2026, Google and the Pixel team co-hosted a showcase at Google I/O India: AI running on Pixel 10’s own silicon. The chip is Google Tensor; inside it is a TPU built for AI work.

One scene in the post: you’re at 30,000 feet with no network and can still hold a deep chat. They also showed models, demo videos, and tools developers can grab now.

Google I/O India joint Tensor and Pixel showcase
Joint showcase at Google I/O India. Source: Google Developers Blog, IMG_5052
✘ Common old path

Ask a question, snap a photo, or transcribe audio—and the phone sends it to the cloud. Offline fails. Online still ships voice and photos off-device.

✔ What this demo shows

The same jobs try to finish on the phone TPU. Model and compute stay local; data does not leave the device.

What shipped

Gemma 4
E2B light build, compiled to run on Pixel TPU
8 clips
Official demos: skills, phone control, vision/audio, industry
100+
Classical models in Model Garden, plus small-language-model packs

These capability types show up again and again in the demos:

Chat Multi-turn talk with no network, including a plane-cabin scene. On the phone TPU
Vision Photograph objects, plants, or a problem on site—no internet required. On the phone TPU
Transcription Lecture or note audio becomes text on-device. On the phone TPU
Settings Voice or text can toggle Wi‑Fi, open maps, and similar actions. On the phone TPU

All four share the phone’s TPU, stay offline where possible, and keep data on the device. Supported devices in the footnote: Pixel 10, 10 Pro, 10 Pro XL, 10 Pro Fold.

How the light Gemma 4 build hits the TPU

The model is Gemma 4 E2B for TPU, already compiled for Pixel TPU—a light build meant to run natively on the phone. Precompiled packs are on the Hugging Face LiteRT community (filenames include Google_Tensor_G5).

If you’re curious about Gemma 4 itself—training, size, and how “thinking” works—there’s a technical-report write-up on this site:

Related reading
Gemma 4 technical report: small params, thinking, and memory efficiency vs large models
Then come back for what it can do offline on Pixel 10.

Eight official demos, by job

Google released eight demo clips. Grouped by job:

1. Offline skills: travel, recipes, home

In the demo, the phone plans trips, suggests recipes, and runs home automation offline. The product name is Agent Skills.

Agent Skills. Travel / recipe / home helper demo. Source: Google I/O India official demo

2. Control the phone: speak or type system actions

Mobile Actions use Functional Gemma (~270M, fine-tuned for function calling). It runs on the TPU so voice or text can drive Wi‑Fi, maps, and similar features with executable actions.

Mobile Actions · Functional Gemma. Voice / text for Wi‑Fi and maps. Source: Google I/O India official demo

3. Chat, see, hear—offline

Four more clips cover chat, snap-and-ask, on-device transcription, and TPU-side chat:

ChatOffline chat
AI Chat: multi-turn offline talk (includes a 30,000-foot cabin scene).
ImageSnap and ask
Ask Image: identify objects, plants, or on-site issues with no internet.
AudioOn-device transcription
Ask Audio: lecture/note audio to text on-device.
TPUChip-side chat
Edge TPU Chat: conversation inferred on the edge TPU.

4. Industry samples: retail floor and auto shop

Two more demos sit in workplaces—still offline:

RetailRecipe → in-store map
Retail: turn a recipe idea into an in-store shopping map, fully offline.
AutoPart photo diagnostics
Auto repair: photograph a faulty part for visual diagnostics.

How developers get a model onto Pixel

The toolkit is Tensor SDK (Beta—apply first). The product page shows three steps:

1 · Compile LiteRT ahead-of-time for supported chips 2 · Deploy via Google Play to Android phones 3 · Infer on Google Tensor TPU Source: developers.google.com/edge/tensor-sdk
Plus Model Garden (100+ ready models) and precompiled small models on the LiteRT community hub.
  1. Apply for Tensor SDK Beta: form on the Tensor SDK page. For commercial redistribution, read the license and distribution terms.
  2. Use stock models or compile your own: Model Garden; small LMs in the Hugging Face Google Tensor collection.
  3. Copy the sample app: GitHub sample_app_tpu and the Tensor SDK Codelab.
  4. Try Gallery App first: Play Store LiteRT AI Edge Gallery. Device scope 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.