On Oct. 6, Google DeepMind released EmbeddingGemma 2, an open model that turns text, code, images, video and audio into numerical representations for on-device search. A text query can be compared with media in the same collection, rather than requiring a separate embedding model for each type.
An embedding model does not search a library by itself. It converts a query and stored items into vectors, or lists of numbers that software can compare for semantic similarity; an app still needs an index and retrieval layer. Google says EmbeddingGemma 2 places all five input types in a shared 768-dimensional space, so developers can build cross-media retrieval on local hardware.
The model has 740 million parameters with all components loaded: a 270-million-parameter text model, plus optional 170-million-parameter vision and 300-million-parameter audio encoders. Google reports that a quantized version used about 191 MB of active RAM in text-only mode and 567 MB with all modalities on a Pixel 11 Pro. Those are company figures for one device, not independent performance measurements.
On MTEB Code v1, using the full-precision, 768-dimensional models, EmbeddingGemma 2 scored 78.68 versus 68.76 for EmbeddingGemma 1. For retrieval tasks, NDCG@10 rewards placing relevant code near the top of the first 10 results; the published Mean(Task) score averages those task results. A higher score means better ranking on that benchmark, not app-level accuracy. The model card lists no results for the earlier model on image, video or audio benchmarks, so the code comparison does not establish better cross-media search.
The weights are available on Hugging Face under Apache 2.0. Google’s model card says developers can shorten vectors to save storage, but quality falls substantially at 128 dimensions, especially for multimodal work. It also warns that performance may vary across the more than 100 supported languages. The release materials provide no independent, application-level test of retrieval quality.
Sources
- Google DeepMind’s launch announcement
- EmbeddingGemma 2 model card and benchmark results
- Model weights and Apache 2.0 license
- MTEB retrieval benchmark and NDCG@10 metric definition
- Image: Google AI for Developers’ model-card banner, used under CC BY 4.0.