EmbeddingGemma 2 brings multimodal search on-device

Google announced EmbeddingGemma 2, an embedding model that places different kinds of content in a shared space for search and retrieval. The company says the open model is based on Gemma 4 and has 740 million parameters.

In practice, a local app could find a video clip from a text query or connect audio and images without sending all data to a cloud service. Google also describes on-device retrieval-augmented generation (RAG) uses.

Before using it, test retrieval quality on your own data and check memory needs, licensing and setup in the model documentation. The published performance figures come from Google’s evaluations.