The Keynote #

Android CLI is great for cheaper models
Android Prototype lets you run apps quickly in a separate lightweight process
New Studio feature to migrate XML to Compose
New set of libraries to manage app state. Making it easier for agents
A2UI lets agents create their own UI
New Android MCP
GoogleBooks
First Android XR Device. XReal Aura with a live demo of the product.

The latest in Android dev tools #

New Devtools

A lot of new android CLI features
Rent remote devices for testing from cli 😢
Run apps directly on Linux using prototype 🤯
Agents tab in studio caught up with the rest of the world

Integrate your app with Android Intelligence #

Android MCP#

Sample App - Jetpacker

Android MCP

Allows Gemini to execute tasks in your apps. Like app intents in iOS
Expose app data as context to Gemini
New skill available with android skills add appfunctions
Shipping this requires registering your MCP in Play Console??
This is an early access feature, you need to apply for access through the QR code at the end of the video

Inside Gemini Nano and Gemma 4 #

What the new Gemma models are capable of On Device#

Offline models only available on some devices. On others, it uses a cloud fallback
Every Model has its own architecture optimized for its use case
Per Layer Embeddings(PLE) way to optimize model memory requirement. Storing in Flash not RAM
Multi Token Prediction(MTP) Uses a draft model to speed up inference upto 3x and 50% wait time reduction. It speculates the output of the main model and shares the KV cache with it. Fascinating concepts, I will suggest you to look these up
Exposed in Android through AI Core
Cool demo of the offline capabilities in the video at 24:10

Optimizing with MLKit APIs #

A lot of cool breakthroughs in running models offline. Gemini Nano 4 uses significantly less RAM than before
MLKit APIs already apply lots of optimizations to make AI integration seamless. It’s still closed source tho
Gemini Nano 4 (2B) is on par with last year’s Gemma 3 (27B) model!!

Use Cases#

Speech Translations - 20+ languages supported
Embeddings - Search app content by meaning
Text to Speech - Human like voices from text
Function Calling - In-App Agents

The rest of the talk demonstrates a use case for MLKit APIs. A fraud detection system

NOTE
Only a limited list of devices support these features. See the list here

Scale your app for every screen with Compose #

A lot of explained examples on how to use the MediaQuery API

NOTE
Here “Media” does not mean audio/video. It means the physical medium through which the user is interacting with

It can be easily used in Previews too!!
Explained examples using the Grids API for UI Layout
Both Grids and FlexBox APIs are now Stable!
Explanation on how the Nav3 Scenes and SceneStratergies work
End segment dives into the engineering ROI of apps with adaptive screens.

NOTE
This talk is basically a Navigation 3 Overview

Google Play - Migration to Compose #

Fascinating approach they took. Crawl, Walk, Jog, Run

Crawl : Converting low traffic screens to compose. Like settings, etc. Goal of developing familiarity with compose
Walk : Converting High traffic components to compose like app bars to understand backwards compatibility and impact to latency and jank.
Jog : High Traffic Pages - fragments within navigation library. Improving latency and jank
Run : Compose navigation and Nav 3. Enabling compose’s shared element transitions

Pre compose theming was tedious. Compose theming is easy to use but there are some performance considerations that I never thought about. The cost of creating a color palette is non-negligible.

Overall compose makes the code more readable, compact, reusable and debuggable than xml.

Cool thing they shared is where they faced problems while using LLMs and what they did to fix them

They shared some tips to improve UX quality

Have design templates bundled with color and typography tokens with built in support for light/dark and contrast modes
Flag hardcoded tokens in the export pipeline
Train the agent in the component catalog to prevent reinventing patterns

rest are standard architecture and code quality practices with compose