Google DeepMind launches EmbeddingGemma 2

Google says EmbeddingGemma 2 maps text, code, images, video and audio into a shared 768-dimensional space.[2] Its listed configurations range from 270 million parameters for text and code to 740 million for all modalities; the other options are 440 million for text and vision, and 570 million for text and audio.[2] The model is based on Gemma 4, has fewer than 1 billion parameters and is released under Apache 2.0.[2] Google calls it “best-in-class for its size,” a company characterization rather than an independently established result in the available information.[1]

Developers can selectively load only the encoders they need at runtime, according to Google.[2] That means a project can choose among the listed modality configurations rather than defaulting to the full multimodal model.[2] Google gives the full model’s size as about 567 MB on a Google Pixel 11 Pro, while its text-only weights can run with about 191 MB of active RAM.[1] Google also says Matryoshka Representation Learning can reduce embeddings from 768 dimensions to 128; at 256 dimensions, image, video and speech retrieval retains about 95% of original quality.[2]

Google says the model is designed for local, privacy-first applications.[1] For retrieval, the company says using it can reduce latency and memory overhead compared with chaining separate image-captioning, speech-to-text and text-embedding models.[1] Google also says the model can route intents in milliseconds without training data or fine-tuning.[1] Google AI Edge Gallery on mobile and Google AI Edge Foresight on Mac are named in the announcement.[1] The company plans to make EmbeddingGemma 2 available as an Android service through ML Kit in the coming weeks, using NPU acceleration to optimize performance across a broad range of devices.[1]