EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind ·
Google DeepMind has launched EmbeddingGemma 2, an open-source multimodal embedding model designed for efficient on-device processing. Built on the Gemma 4 architecture with 740 million parameters, it natively maps text, images, audio, and video into a unified embedding space. The model achieves leading performance among sub-1B models, significantly improving code understanding while maintaining high quality for vision and audio tasks. It features modular design options, allowing developers to use smaller parameter counts for specific modalities like text-only workloads. Storage efficiency is enhanced through Matryoshka Representation Learning, which enables dynamic vector truncation for reduced memory usage. Optimized for edge hardware, the model runs on devices such as the Google Pixel 11 Pro with minimal RAM requirements. With an 8K token context window, it can process substantial amounts of media directly on local hardware without cloud dependency. This release supports privacy-first applications by enabling offline semantic search and retrieval-augmented generation pipelines.
deepmind