Google Releases EmbeddingGemma 2 for Local Multimodal Search

Google DeepMind released EmbeddingGemma 2, a 740 million parameter open weight model for local multimodal search, as detailed in a Google developer blog post. It maps text, images, video frames and audio into one vector space, allowing applications to search and classify local content without separate captioning, transcription and text embedding models.

The model uses about 191 MB of active memory with text weights and about 567 MB with all modalities on a Google Pixel 11 Pro. Its modular design lets developers load only the required encoders. It also supports classification against labels and descriptions without training or fine tuning.

Google added Instant Media Search and Video Moments Finder demonstrations to the AI Edge Gallery app. These tools search local photos and videos through text, images or camera input, with embeddings and similarity ranking processed on the device. The company also released AI Edge Foresight for Mac, which searches meeting transcripts, notes, documents and images using local processing.

MediaPipe Tasks supports the model through its Universal Embedder, Semantic Retriever and Decision tasks. Google plans to add access through ML Kit on Android in the coming weeks, including NPU acceleration where available. Developers can also run the model through LiteRT across Android, iOS, macOS, Windows, Linux and the web.