We launched EmbeddingGemma final 12 months to supply a light-weight possibility for high-quality textual content embeddings, to assist your apps arrange, search, and join data straight on shopper {hardware}. The developer group’s response blew previous our expectations. With greater than 20 million downloads, builders have used it to energy smarter on-device search instruments and privacy-first retrieval augmented technology (RAG) pipelines.
In the present day, we’re launching EmbeddingGemma 2, increasing past textual content to unify code, pictures, video, and audio in a shared embedding house. Constructed on the Gemma 4 structure and launched beneath a commercially permissive Apache 2.0 license, EmbeddingGemma 2 has 740 million parameters, making it optimum for on-device inference. It may assist discover a particular video clip from a voice memo, or search by hours of audio recordings primarily based on a textual content question, all processed by a single, natively multimodal mannequin.
