Google Releases EmbeddingGemma 2 for Natively Multimodal Embeddings

Google has announced EmbeddingGemma 2, an open-source 740-million parameter model designed for natively multimodal embeddings. Built on the Gemma 4 architecture, it maps text, image, audio, and video inputs into a shared vector space. By providing a state-of-the-art open-weights model for unified embeddings, EmbeddingGemma 2 significantly advances multimodal semantic search and Retrieval-Augmented Generation (RAG). It enables developers to perform cross-modal search and indexing across diverse media types without relying on proprietary cloud services. EmbeddingGemma 2 projects all supported modalities into a standardized 768-dimensional vector space. With a lightweight footprint of 740 million parameters, the model is optimized for high-performance cross-modal retrieval and on-device or edge deployment.