EmbeddingGemma 2

Google DeepMind has launched EmbeddingGemma 2, a lightweight 740-million-parameter multimodal embedding model that can process text, images, audio, and video in a shared embedding space for on-device inference. The model achieves best-in-class performance for its size, requires minimal memory (as low as 191MB for text-only on a Pixel 11 Pro), and supports an 8K token context window, enabling privacy-first applications like local semantic search and retrieval-augmented generation. Released under an Apache 2.0 license and built on Gemma 4 architecture, it represents a significant expansion from the original EmbeddingGemma, which achieved over 20 million downloads for text-only embeddings.