Skip to content
Predictive Systems
PSI Daily

Google Shrinks Smart Search to Fit in Your Pocket

Privacy

Google's new EmbeddingGemma 2 lets your phone search photos, notes and voice memos by meaning without sending them to the cloud. Small models just made privacy practical.

A phone on a dark desk with small photo and document tiles floating above it, one glowing amber, beside the EmbeddingGemma 2 logo.

The smartest search on your phone no longer has to send your life to the cloud. Google's new EmbeddingGemma 2, released October 6, is a small, free model built to find photos, notes, voice memos and videos by meaning, right on the device.

An embedding turns a photo, sentence or sound into a list of numbers, so things with similar meanings land close together. That is how an app can find "the beach trip with Grandma" even when no file carries those words. It is also how AI assistants pull up the right notes before they answer, a method known as RAG.

Light Enough to Run Anywhere

EmbeddingGemma 2 has 740 million parameters, the settings a model learns, and handles text, code, images, audio and video in one shared space. Google says the full version needs about 567MB of memory on a Pixel 11 Pro, and a text-only version about 191MB. Developers can load only the parts they need.

It ships under the Apache 2.0 license, so anyone can use it commercially, and the weights are on Hugging Face and Kaggle. Google says the first version, which handled only text, passed 20 million downloads, and that version two scores nearly 10 points higher on a code search test. Those benchmark results are Google's own and have not been independently checked.

Keep the Memory at Home

The bigger story is where your data lives. When the indexing happens on your device, your photos, voice notes and documents do not have to be uploaded to a server just to become searchable, and Google says it can work fully offline. That is a quiet win for privacy, much like the on-device agent from Liquid AI we covered earlier.

For builders, the takeaway is practical. Before you ship a feature that searches people's private files, test whether a small local model can do the job. If it can, the safest data is the data you never collect.

Sources

  1. Google, The Keyword, EmbeddingGemma 2: an open, lightweight multimodal embedding model, October 6, 2026
  2. Hugging Face, google/embeddinggemma-2 model card, accessed October 9, 2026
  3. Google Developers Blog, EmbeddingGemma 2: The Developer Guide, October 6, 2026
  4. Google Developers Blog, Bring multimodal semantic search to the edge with EmbeddingGemma 2, October 6, 2026
  5. Crypto Briefing, Google releases EmbeddingGemma 2, a 740 million parameter open-weight model under Apache 2.0, October 6, 2026
  6. Predictive Systems, Liquid AI's LFM2.5-2.6B Runs AI Agents On-Device, August 2026