EmbeddingGemma 2 is an open multimodal embedding model by Google DeepMind (huggingface.co)

Google DeepMind has announced EmbeddingGemma 2, an open multimodal embedding model that integrates text, images, video, and audio into a single 768-dimensional vector space. With 740 million parameters, the model features a 270M parameter text backbone and separate encoders for vision (170M) and audio (300M), making it suitable for deployment on consumer hardware like mobile devices. EmbeddingGemma 2 offers impressive low-latency semantic representations ideal for applications such as search, retrieval-augmented generation, classification, and clustering. This model builds upon advancements from its predecessor and introduces several key features: native multimodality unifying four data types, improved multilingual support across 100+ languages, and significant enhancements in coding tasks—showing a ~14% performance improvement. Additionally, the innovative Matryoshka Representation Learning allows for flexible vector truncation to reduce storage costs while maintaining quality. With an 8K token context window and the ability to process multi-modal inputs, EmbeddingGemma 2 represents a significant leap in making sophisticated AI capabilities more accessible and efficient for developers, paving the way for a wider range of applications in AI/ML.