740M parameter model runs on mobile, Google’s Embedding Gemma 2 enables one-stop search for text, images, audio, and video.

Beating AI Express: Google DeepMind has open-sourced EmbeddingGemma 2. The model has just 740 million parameters, capable of directly processing text, code, images, video, and audio, and embedding all of them into a shared vector space for search and Retrieval-Augmented Generation (RAG). The previous-generation EmbeddingGemma only supported text processing. This new model allows partial loading as needed: 270 million parameters for text and code only, 440 million when adding vision capabilities, and a full 740 million parameters when all components are loaded. Google tested the model on the Pixel 11 Pro; the quantized text-only version uses as little as ~191MB of active memory, while the full version takes around 567MB. Its context window has been expanded from 2K in the prior generation to 8K, supporting processing of up to ~5.5 minutes of audio, 29 images, or 58 video frames in a single pass. The most notable improvement is in code retrieval: its MTEB Code score rose from 68.76 in the previous version to 78.68, while multilingual text performance remained largely stable, increasing slightly from 61.15 to 61.36. The model also supports reducing the default 768-dimensional vectors to 512, 256, or 128 dimensions, cutting vector storage requirements by up to one-sixth. The license has also been relaxed: while the prior EmbeddingGemma used Google’s proprietary Gemma terms, EmbeddingGemma 2 is released under the Apache 2.0 license. Gemma 4, launched by Google earlier this year, also uses Apache 2.0, reflecting a clear trend of recent open models being more business-friendly and accessible for secondary development.