Google launches EmbeddingGemma 2, a multimodal model that runs offline on devices

Google's lightweight embedding model EmbeddingGemma has been downloaded more than 20 million times since last year. It is widely used in on-device search tools and privacy-first RAG pipelines.

Now Google has released EmbeddingGemma 2, built on the Gemma 4 architecture with 740 million parameters under the Apache 2.0 license. It moves beyond text to handle code, images, video and audio in one shared embedding space.

The model has a modular design, needing only 270 million parameters for text tasks. Vector storage can shrink up to 6 times. Code scores rose from 68.76 to 78.68 on the MTEB Code benchmark. On a Pixel 11 Pro, the text-only weights need about 191MB of memory.

The weights are already available on Hugging Face and Kaggle. Local embedding generation keeps data private and cuts delays, letting developers build fully offline cross-modal search systems.