Google publishes technical report for EmbeddingGemma 2 multimodal embedding model

Google publishes the technical report for EmbeddingGemma 2, a 740M-parameter open multimodal embedding model.

The report covers the model's architecture, training recipe, and ablations. Google says the model places text, code, images, audio, and video in one shared embedding space.

Google says EmbeddingGemma 2 outperforms specialist and general models up to 12 times larger. For on-device use, the report lists modular encoders, Matryoshka dimensions, and int4 quantization that reduces the model to 567MB.