Google is launching EmbeddingGemma 2 today, bringing audio, video, images and code into a shared embedding space for apps that search and connect information. The model has 740 million parameters and uses the Gemma 4 architecture, with Google positioning it for inference on consumer devices. Its predecessor handled text embeddings, making support for other media the central change in this release.
EmbeddingGemma 2 is released under Apache 2.0, a license that permits commercial use. Google says a voice memo could serve as the input for finding a particular video clip, while a written query could retrieve material from hours of recorded audio. The company says one model processes those different forms of media.
Read more: Google Fined €403 Million Over Location Tracking After BEUC Complaint
Google introduced the original EmbeddingGemma last year, aiming to let apps organize information, connect it and perform searches on consumer hardware. That model has recorded more than 20 million downloads, according to Google. Developers have used it for search on devices and retrieval augmented generation (RAG) systems designed around privacy, the company says.
Google previously expanded the inputs accepted by its Search field in May with Gemini 3.5 Flash, allowing conversational queries alongside material such as images and Chrome tabs. That earlier overhaul also introduced background Search agents for monitoring topics, with an initial summer launch planned for U.S. AI Pro and Ultra subscribers.
