SiliconANGLE reports that it is based on the Gemma 4 architecture and has 740 million parameters, while its 270-million-parameter text core used about 191 MB in a quantized test on a Google Pixel 11 Pro. Embeddings contain 768 values and can be reduced to 128; Google says retrieval quality for image, video and speech remains about 95% at 256 dimensions. The weights are available on Hugging Face and Kaggle under Apache 2.0, and work with vLLM, llama.cpp and Ollama. The release extends the earlier text-only EmbeddingGemma line to multimodal retrieval, including on-device use cases such as finding a video moment from a voice query.
Raw SignalOct 10, 20261 min read1 source
Google expands EmbeddingGemma beyond text to images, audio and video
In one minute
Google has released EmbeddingGemma 2, an open multimodal embedding model that maps text, images, audio and video into a shared embedding space.