EmbeddingGemma 2 puts text, image, audio and video in one search space

Google released EmbeddingGemma 2 on 6 October, expanding its small embedding model beyond text. The Apache 2.0 model maps code, images, video and audio into a shared representation, allowing applications to retrieve one medium using a query in another. The full model has 740 million parameters.

A text-only configuration uses 270 million parameters, with optional vision and audio encoders. Developers can also shorten output vectors from 768 to 128 dimensions to reduce storage, accepting a representation trade-off. Google specifies an 8K-token context window. These are practical controls for a local search index: a phone or laptop does not need the same configuration as a larger server. Model benchmarks still need to be checked against the actual collection being searched.