Google has released EmbeddingGemma 2, a 400-million-parameter open model that turns text and images into vectors for search, retrieval and similarity matching. The model is designed for applications that need to index mixed media without sending content to a large hosted system.
Google says EmbeddingGemma 2 outperforms some competing embedding models with more than twice its parameter count. That comparison is based on the company’s reported benchmarks and should be checked against the languages, image types and retrieval tasks in a real deployment. A smaller parameter count can reduce memory and compute requirements, but it does not guarantee lower latency on every device or data set.
Multimodal embeddings place related text and images in a shared numeric space, allowing a text query to retrieve an image or an image to find related documents. The open release gives developers the option to run and adapt that capability locally. Its practical value will depend less on a single aggregate score than on recall, ranking quality and resource use with an organization’s own material.