The company says it is built for on-device inference and released under Apache 2.0. Google lists an 8K-token context window, modular encoders, and vector truncation from 768 dimensions to 128. It reports a 9.92-point improvement over EmbeddingGemma on MTEB Code. These benchmark and performance statements are Google’s claims; the announcement does not provide independent validation. The model is available via Hugging Face and Kaggle, with deployment paths including MediaPipe and LiteRT.
Raw SignalOct 7, 20261 min read1 source
EmbeddingGemma 2 is a best-in-class open model for natively multimodal embeddings
In one minute
Google announced EmbeddingGemma 2, an open-weight embedding model with 740 million parameters that maps text, code, images, video, and audio into a shared embedding space.