EmbeddingGemma 2 puts multimodal search embeddings on-device
Google’s open-weight EmbeddingGemma 2 maps text, code, images, video, and audio into one 768-dimensional space, enabling a local search system to retrieve across media types without separate embedding models. InfoQ reports that the model scales from 270 million parameters for text and code to 740 million for full multimodal use and is designed for phones and laptops.
The model also supports truncating vectors to reduce storage: the article gives a sixfold reduction at 128 dimensions, with Google-reported quality retention of about 90% for text but 75% for image, video, and speech retrieval. InfoQ quotes a Reddit commenter who sees cross-modal search as the main advance and questions how much text-only performance improved; InfoQ has no reader-comment section of its own.