AI
Google launches EmbeddingGemma 2, a multimodal AI model for on-device semantic search
Google DeepMind has released EmbeddingGemma 2, an open-weight embedding model that runs locally on phones and computers to power text, image, video, and audio search without internet or privacy risks.
Google Developers Blog (AI)
Google DeepMind has unveiled EmbeddingGemma 2, a compact multimodal embedding model designed to bring sophisticated search capabilities directly to users' devices. The 740-million-parameter model maps text, images, video frames, and audio into a single vector space, enabling developers to build fast, privacy-preserving search and retrieval features without the overhead of chaining together separate transcription, captioning, and embedding systems.
What makes it efficient#
The model achieves remarkable efficiency for an edge device. It can run on as little as 191MB of RAM for text-only operations and 567MB for the full multimodal version—small enough to function smoothly on a Google Pixel 11 Pro. Using quantization techniques, the model compresses its weights to INT4 and INT8 formats, bringing advanced vector search within reach of mid-range consumer hardware. Its modular design means developers can load only the encoders they need, reducing memory pressure further.
Try it out now#
Google is releasing Instant Media Search and Video Moments Finder as interactive demos in the Google AI Edge Gallery app, available on Android and iOS.
Instant Media Search lets users find photos and videos in their device libraries using natural language queries or example images. As they type—searching first for "cat," then "cat sleeping on keyboard"—the app instantly re-ranks results and highlights matching content. Visual similarity matching works too: users can select any photo or point a camera at an object to find visually related images, all without leaving the device.
Video Moments Finder demonstrates the cross-modal power of EmbeddingGemma 2 by letting users search local video files using plain English. Typing queries like "kids laughing" or "dog catching a frisbee" instantly locates and marks the exact timestamps where those moments appear in the video.
A second demo app, AI Edge Foresight, brings these capabilities to Mac users in a meeting-focused tool. It uses on-device AI to enrich note-taking in real time, answer questions during calls, and search across personal files—all while keeping sensitive data local and secure.
