Google launches EmbeddingGemma 2, a 740M open multimodal embedding model for on-device text, image, audio and video retrieval

What changed

Google DeepMind launched EmbeddingGemma 2 on October 6, extending the earlier text-only embedding model into a native multimodal embedding space spanning text, code, images, audio and video. Built on the Gemma 4 architecture, it has about 740 million parameters and is released under Apache 2.0, targeting on-device semantic search, private RAG and low-latency retrieval. Google positions it for consumer hardware; benchmark claims should still be separated from later independent evaluations.