🔍Google Releases EmbeddingGemma 2, a 740M Multimodal Embedder for On-Device Use

Your local search just got video and audio support

Google just shipped a multimodal embedding model small enough to run on your laptop, and it handles text, images, audio, and video in one vector space. EmbeddingGemma 2 is open, has 740 million parameters, and is available now on Hugging Face and Kaggle. Gemini Enterprise Agent Platform availability is coming soon. If you're building search, RAG, or routing features, this removes the usual trade-off. You can index a voice memo and jump straight to the matching video clip, all locally. Privacy-sensitive apps and offline edge deployments get cross-modal retrieval without shipping data to an API. Pair it with a Gemma 4 generator and you have an on-device RAG pipeline that understands more than text. The numbers are the story. Text-only weights need about 191MB of active RAM, and the full multimodal model needs about 567MB. The 8K context window covers roughly 5.5 minutes of audio or 58 video frames. Vectors shrink from 768 dimensions to 128 through Matryoshka truncation, and code retrieval jumped 9.92 points on MTEB Code, from 68.76 to 78.68.