Google's EmbeddingGemma 2 can represent text, code, images, video and audio in a shared embedding space. The model is designed for semantic search and retrieval tasks on consumer devices, including phones and laptops.
Each representation has 768 dimensions, so developers can compare content from different media types by semantic similarity. The model can support searches for documents, screenshots and recordings, as well as natural-language searches of source code and text-based searches of video.
Developers can load the encoders their applications need. The text and code configuration has 270 million parameters, while full multimodal support uses 740 million. Google says the model significantly improves on EmbeddingGemma 1 for code search and technical retrieval.
Read nextTwo German courts hold Google responsible for false AI search claimsShortening vectors reduces storage needs
EmbeddingGemma 2 also supports Matryoshka Representation Learning, which lets developers shorten its vectors to reduce storage needs. At bfloat16 precision, one million vectors with 768 dimensions take about 1.5 GB. Shortening them to 128 dimensions cuts that figure to about 250 MB.
That saving comes with a retrieval quality trade-off. At 128 dimensions, the model retains approximately 90% of quality for text retrieval, compared with around 75% for image, video and speech retrieval. Google recommends truncation only when storage needs or search speed call for it.
Reddit users commented on the announcement
In comments on the announcement, Reddit user Distinct-Target7503 welcomed the shared space for images and audio but described the text-only improvement as modest. Reddit user Open-Adhesiveness-86 pointed to lessons from the first EmbeddingGemma about using task prefixes correctly and normalizing vectors after truncation, while saying those lessons probably still apply.
The original EmbeddingGemma was introduced last year with a focus on offline and privacy-sensitive uses, including local file search. EmbeddingGemma 2 is available on Hugging Face.
