EmbeddingGemma 2 (developers.googleblog.com)

The recent release of EmbeddingGemma 2 marks a significant advancement in multimodal AI models, particularly for retrieval augmented generation (RAG) applications. This new model, which is compact yet powerful with less than 1 billion parameters, is tailored for diverse content types including text, code, images, video, and audio. It leverages a unique modular architecture that enables users to load only the necessary encoders, scaling from a lightweight 270M parameters for text and code to 740M for full multimodal capabilities. This flexibility allows for high retrieval accuracy with low latency and minimal computational requirements, making it accessible for a wider range of applications. EmbeddingGemma 2 excels in generating embeddings that occupy a shared 768-dimensional space across all modalities, facilitating seamless comparisons of different content types. It significantly enhances previous capabilities, boasting a 14% increase in performance on multilingual text tasks while introducing retrieval functionalities for images, video, and audio. Its efficient memory management allows for easy scaling—users can truncate embeddings and omit unused encoders at load time. This innovation positions EmbeddingGemma 2 as a versatile tool for developers and researchers in the AI/ML community, paving the way for more integrated and efficient multimodal applications.