The 740 million parameter EmbeddingGemma 2 converts multimodal data into vectors using only 191 MB of RAM. Google claims it outperforms rival embedding models twice its size. This efficiency enables offline RAG applications when paired with Gemma 4.