Google DeepMind has introduced Embedding Gemma 2, an open, lightweight model that natively maps text, images, audio, and video into a unified embedding space. It follows the original Embedding Gemma, which surpassed 20 million downloads as developers used it for on-device search and privacy-first retrieval augmented generation (RAG) pipelines.
Embedding Gemma 2 is built on the Gemma 4 architecture, is released under a commercially permissive Apache 2.0 license, and has 740 million parameters, making it suited to on-device inference. It can locate a specific video clip from a voice memo or search hours of audio recordings using a text query, all handled by a single model. The company says it achieves leading scores among sub-1B multimodal embedders on MTEB Code and MAEB benchmarks.
The model is modular, needing as little as 270M parameters for text-only workloads, with optional vision and audio encoders. Using Matryoshka Representation Learning, developers can truncate output vectors from 768 dimensions down to 128, cutting storage needs by up to 6x. With quantization, it requires about 191MB of active RAM for text-only weights on a Google Pixel 11 Pro, and about 567MB for the full multimodal model.
Embedding Gemma 2 also offers an 8K token context window, four times larger than its predecessor, and shows a 9.92-point gain on code benchmarks. Because it shares its tokenizer and audio encoder with Gemma 4, developers can run both models together in a unified pipeline with a lower combined memory footprint.
Source: Google DeepMind Blog · Summarized by HeadlinesBriefing