Stories
Google Launches EmbeddingGemma 2 For On-Device Multimodal Search
- •Google has launched EmbeddingGemma 2, an open natively multimodal embedding model designed to run efficiently on devices.
- •Based on Google's Gemma 4 architecture, the model is released under the commercially permissive Apache 2.0 licence.
- •The model is designed to handle text, code, images, video and audio within the same system, enabling cross-modal searches.
- •It supports offline and privacy-focused retrieval-augmented generation for local information processing without cloud servers.
- •The modular design allows developers to use only needed components and supports Matryoshka Representation Learning to save storage space.
Key Questions Answered
EmbeddingGemma 2 is Google's new open multimodal embedding model designed to run efficiently on devices, converting information like text, images, audio, and video into numerical vectors. The model is based on Google's Gemma 4 architecture and is released under the commercially permissive Apache 2.0 licence.
The full model contains 740 million parameters, but its modular design allows developers to load only the parts they need. The text and code version uses 270 million parameters, adding vision increases it to around 440 million, incorporating audio brings it to roughly 570 million, and the complete multimodal version reaches 740 million.
EmbeddingGemma 2 supports Matryoshka Representation Learning, allowing its 768-dimensional vectors to be reduced down to as little as 128 dimensions to save storage space. Using 256 dimensions retains most of the original search quality for text and code, and about 95% of the quality for image, video, and speech retrieval.
