Google EmbeddingGemma 2 Brings Multimodal AI Search On Device

Summary

On October 6, 2026, Google DeepMind engineers Sahil Dua and Henrique Schechter Vera launched EmbeddingGemma 2, a commercially permissive Apache 2.0 model with 740 million parameters, built on Gemma 4 and Gemini Embedding technology. It maps text, code, images, video and audio into one space for on-device search, retrieval, routing and RAG. Its text-only predecessor exceeded 20 million downloads.

Google says it leads sub-1B multimodal embedders on MTEB Code and MAEB, matches EmbeddingGemma’s multilingual text quality, and beats some specialist models over twice its size. MTEB Code rose 9.92 points, from 68.76 to 78.68. The modular design uses 270M parameters for text, plus optional 170M vision and 300M audio encoders. MRL shortens 768-dimension vectors to 512, 256 or 128, cutting local vector database storage and memory by up to 6x. Quantized Pixel 11 Pro use is about 191MB active RAM for text weights and 567MB for the full model. Its 8K context, 4x EmbeddingGemma 1, handles 5.5 minutes of audio, 29 images, 58 video frames or mixed inputs.

Local generation keeps data private, lowers latency and works offline. Pairing it with Gemma 4 creates multimodal on-device RAG with lower combined memory because both share a tokenizer and audio encoder. Google AI Edge Gallery offers Instant Media Search and Video Moments Finder; AI Edge Foresight combines retrieval with Gemma 4 reasoning; MediaPipe Decision Task API supports real-time classification, routing and prediction. Weights are on Hugging Face and Kaggle, with LiteRT Community builds and Gemini Enterprise Agent Platform Model Garden access coming soon. Deployment supports Google AI Edge MediaPipe, LiteRT, transformers.js, WebGPU, transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio; Qdrant stores vectors, while Unsloth provides fine-tuning guidance.