Google's EmbeddingGemma 2 Beats Larger Rivals

Google released EmbeddingGemma 2 this week, an open embedding model that turns text, images, video, audio, and code into numerical vectors, the format machines use to compare and search content. At 740 million parameters, the company calls it the most compact model of its kind on the market, and claims it beats rival embedding models up to twice its size on multimodal benchmarks.

The model is built to run locally, with no API key required. Google says a query takes between 20 and 70 milliseconds using WebGPU in a browser, and the whole thing fits in roughly 191 MB of RAM. It also shrinks local vector database storage by as much as six times compared to other options. Developers working only with text can skip the full model and use a smaller 270-million-parameter version instead.

Google is positioning EmbeddingGemma 2 as a companion to its small open language model, Gemma 4, for building retrieval-augmented generation apps that work entirely offline, without routing data through external servers. The weights are already posted on Hugging Face and Kaggle, alongside a developer guide. Here's what Google is saying about the performance claims.