Google’s EmbeddingGemma 2 can represent text, source code, images, video, and audio in one 768-dimensional embedding space, according to InfoQ. That gives developers a common basis for comparing meaning across media instead of maintaining separate, incompatible search indexes for each format.
Embeddings are compact numerical representations used to retrieve related material. With a shared space, an application could use a natural-language query to find a document, screenshot, recording, code fragment, or video segment based on semantic similarity. Google positions the model for search, retrieval-augmented generation, classification, and clustering on consumer hardware such as laptops and phones. Developers can load only the encoders needed by their application rather than every supported modality.
The model is based on Gemma 4 and released under the Apache 2.0 license, making its weights available for local deployment and modification under that license. A unified representation simplifies application architecture, but it does not guarantee equal retrieval quality across every language, media type, or specialized domain. Developers will still need to evaluate recall, ranking, resource use, and privacy for their own data.