How can your phone search your photos, videos and voice notes by meaning — without sending them to the cloud?

Google DeepMind's EmbeddingGemma 2 is a small open model that turns text, images, audio and video into meaning-numbers right on your phone, so you can search your own files by describing them — no cloud upload required.

You take a photo of a receipt in March, record a voice memo in June, and save a clip in September — by October, finding any of them means scrolling and hoping. Short answer: A new open model from Google DeepMind, EmbeddingGemma 2, lets your phone understand what your files are about and match them to a plain-language search, entirely on the device. Nothing has to be sent to the cloud to make it work.

Here is how. An “embedding” is a list of numbers that captures meaning: things with similar meanings get similar numbers, so a photo of a beach at sunset lands close to the words “beach sunset” in number-space, and far from a tax form. EmbeddingGemma 2, released October 6, 2026, maps text, code, images, audio and video into one shared space of 768 numbers, so a single typed question can search across all of them at once.

It is built in pieces — a 270-million-parameter text core, plus optional vision (170M) and audio (300M) encoders you only load if you need them — and it is released under the Apache 2.0 license, so anyone can build with it. Google’s figures for a Pixel 11 Pro put the quantized model at about 191MB of RAM for text-only search, or about 567MB with every encoder switched on. Its vectors also use Matryoshka learning, meaning they can be trimmed from 768 numbers down to 512, 256 or 128 — cutting storage by up to about six times, with a modest quality trade.1.

Install Google AI Edge Gallery. It is Google’s free showcase app for on-device AI models.2. Open the media showcases.

Try Instant Media Search or Video Moments Finder, which use the model to search photos and clips by description.3. Search by meaning, not filename. Type what you remember — “the dog in the snow,” “the receipt from the hardware store” — and let the numbers do the matching.4.

Building your own? The open weights are published on Hugging Face, ready to drop into an app of your own.Pro tip: The privacy benefit is the headline feature, not a footnote. When the index of your photos and voice notes lives on your phone, searching your own life stops meaning uploading your own life first.Ad’s take: Search that understands meaning is how our photo libraries should have worked all along — and doing it on-device makes the cloud copy of your memories optional.