Google’s EmbeddingGemma 2 is an open-weight model for search and retrieval that supports text, code, images, video and audio. Google says it maps them into a shared 768-dimensional space, so developers can search across media types without maintaining separate embedding spaces. The model is available on Hugging Face under the Apache 2.0 license.
Developers can load only the encoders they need: the model ranges from 270 million parameters for text and code to 740 million for full multimodal support. Its vectors can also be shortened to save memory, but Google reports that at 128 dimensions, retrieval quality falls to about 75% for images, video and speech.
Why it matters
Developers building local search can use one model to retrieve across documents and media, but shrinking vectors to save storage may reduce results for non-text searches; the source does not provide independent test results.
