Google has released EmbeddingGemma 2, an open-source AI model designed for local search across text, code, images, audio, and video. It converts data into 768-dimensional numerical vectors, enabling semantic search without sending user data to the cloud.
Built on the Gemma 4 architecture, the model contains 740 million parameters and is licensed under Apache 2.0. It features a modular design: 270 million parameters for text, 170 million for images, and 300 million for audio. It supports Matryoshka Representation Learning to reduce vector dimensions.
In tests on a Pixel 11 Pro, the text-only version required 191 MB of RAM, while the multimodal version used approximately 567 MB. The model outperformed other sub-1 billion parameter models in MTEB and MAEB benchmarks, with its MTEB Code score rising from 68.76 to 78.68.
EmbeddingGemma 2 supports local RAG pipelines when paired with Gemma 4. The model is available on Hugging Face and Kaggle, with support for tools such as transformers, llama.cpp, Ollama, and LM Studio.
