Google DeepMind launches EmbeddingGemma 2, a 740M-parameter multimodal embedding model

Google DeepMind announced EmbeddingGemma 2 on October 6, 2026, describing a model that maps combinations of text, images, audio, and video into one embedding space. Built on the Gemma 4 architecture, it has 740 million parameters and is released under an Apache 2.0 license that the announcement describes as commercially permissive. Google says the model is optimized for on-device inference.

The announcement describes configurations for different workloads: text-only use can require as little as 270 million parameters, while the optional vision and audio encoders are listed at 170 million and 300 million parameters. Example tasks include finding a video clip from a voice memo and searching audio recordings with a text query, bringing those cross-modal retrieval tasks within the same model’s scope.

Google also says Matryoshka Representation Learning lets developers truncate output vectors from 768 dimensions to 512, 256, or 128, with up to a sixfold reduction in storage and memory use. As background, Google says it introduced the previous EmbeddingGemma as a lightweight text-embedding model the year before and that it had passed 20 million downloads. EmbeddingGemma 2 expands that model line into multimodal embeddings.