EmbeddingGemma 2 searches across different media

One space for several media

On 6 October 2026, Google launched EmbeddingGemma 2, released under Apache 2.0. The model represents text, code, images, video and audio in a shared vector space.

Connect a question to a different kind of content

An embedding turns content into numbers so its proximity to other content can be compared. Multimodal retrieval could let a text query find an audio clip or image without requiring a manually written description for every search target.

For a resource library or game project, a trial might combine known screenshots, sounds and documents. The team would define expected results in advance and examine misleading matches too: visual resemblance does not necessarily imply the same domain meaning.

Local does not automatically mean private

Model choice alone cannot describe a complete application. Indexes, logs, telemetry and any external calls need examination before claiming local operation. Retrieval quality should also be measured separately for each media type, alongside indexing memory and duration on the target hardware.

Official source.

Editorial retrospective: event dates are stated in the text.