Google releases EmbeddingGemma 2, an open multimodal embedding model

EmbeddingGemma 2 is based on Gemma 4 and is designed to run on consumer hardware such as phones and laptops, with low-latency output for search, RAG, classification and clustering. Developers can load only the encoders they need, starting at 270 million parameters for text and code and scaling to 740 million for full multimodal support. Google says the model significantly outperforms EmbeddingGemma 1 on code search and technical retrieval.

Because all media types share one embedding space, a single index can match a text query against documents, screenshots, recordings and video. The model supports Matryoshka Representation Learning, so vectors can be shortened to save storage. Google's example puts a million 768-dimensional vectors at bfloat16 at about 1.5 GB, and 128 dimensions at about 250 MB, a sixfold reduction.

Google recommends truncating only when storage or search speed requires it, since at 128 dimensions the model keeps about 90% of quality for text but only around 75% for image, video and speech retrieval. The model is available on Hugging Face.

Agents that search code, documents and media can use one local index instead of maintaining separate embedding spaces for each media type. Truncating vectors trades retrieval quality for storage, and the loss is larger for non-text media. Reddit commenters quoted in the article warned that task prefixes affect recall and that truncated vectors should be sliced before normalizing, so check both before indexing.