Google's New AI Runs on Your Phone, Understands Text, Images, Audio

Google DeepMind just released EmbeddingGemma 2, a new AI model that can understand and compare text, images, audio, and video all at once. Think of it as a universal translator for computers: it converts any of these into a list of numbers (called an embedding) that captures meaning. That lets your phone find a photo by describing it, or search your voice notes by typing a phrase. The best part? It's designed to run entirely on your device, so your data never leaves your phone.

The model is small—740 million parameters, about the size of a few photos—so it works on mobile devices and laptops without an internet connection. It supports over 100 languages and handles code 14% better than its predecessor. Developers can choose which parts to load, saving space. It also uses a trick called Matryoshka Representation Learning that shrinks the number of numbers needed to store each item, cutting storage costs by up to 6x with little quality loss.

For everyday people, this means future apps could offer smarter search, better organization, and more privacy. Imagine asking your phone 'find the video where I talked about vacation plans' and it actually works—without uploading anything. Businesses could use it for customer support, sorting documents, or recommending products, all on local hardware. The catch is that this is a building block for developers, not a finished app. You won't download it directly, but you'll see its benefits in apps later this year.