One search model for many formats
Google DeepMind introduced EmbeddingGemma 2 on October 6. The 740-million-parameter open model embeds text, code, images, audio and video in a shared vector space for semantic retrieval.
A user could describe a scene to find a matching segment in a video library, or search voice recordings using written queries. Unlike a chat assistant, the model's primary role is representing and retrieving related content.
Why on-device inference matters
EmbeddingGemma 2 is built on Gemma 4 technology and released under Apache 2.0. Its modular architecture can run with about 270 million parameters for text-only tasks, with optional vision and audio encoders.
Google reports approximately 191 MB of active RAM for a quantized text-only configuration and 567 MB for the full multimodal system on a Pixel 11 Pro. These are vendor-reported measurements, not universal requirements for every device.
Practical use cases
Developers can build private search across media libraries, local code indexing, video clip retrieval and retrieval-augmented generation systems. A separate generative model can then use results from these local searches as supporting context.
The model produces 768-dimensional vectors and supports truncation to 512, 256 or 128 dimensions, potentially saving storage space. The trade-off should be tested with the actual search corpus.
Limits to watch
Google cites an 8,000-token context window and support for mixed media. Actual performance depends on hardware, indexing strategy, compression and input complexity.
The story is not about replacing cloud frontier models. It is about making cross-media retrieval more practical when privacy, offline use and latency matter.
Official sources and original visuals
The authors' official images and charts are available through the source links. This article's cover is independent editorial artwork, not an official vendor image.