It maps text, code, images, video, and audio into a shared 768-dimensional vector space. The model can be loaded in modular configurations: 270 million parameters for text and code, 440 million for text plus vision, 570 million for text plus audio, and 740 million for all modalities. The guide also describes Matryoshka truncation from 768 to 128 dimensions and says that, in bfloat16, one million vectors require roughly 1.5 GB at 768 dimensions versus 250 MB at 128. Google says the model is available through sentence-transformers 6.1.0 or later. These are product details and storage examples presented by Google; the supplied source does not independently validate comparative retrieval performance.
Raw SignalOct 6, 20261 min read1 source
Google’s developer guide introduces EmbeddingGemma 2, an open embedding model built for multimodal retrieval
In one minute
Google’s developer guide introduces EmbeddingGemma 2, an open embedding model built for multimodal retrieval.