EmbeddingGemma

Google designed EmbeddingGemma for embedding text on the device where the data lives: a 308M-parameter multilingual model built on Gemma 3 that runs in under 200 MB of RAM with quantization and generates embeddings in under 22 ms on EdgeTPU. Matryoshka Representation Learning lets one model emit vectors anywhere from 768 down to 128 dimensions, so applications can trade accuracy against storage without retraining, and a 2K token context window covers most document chunking schemes. Trained on more than 100 languages, it targets information retrieval, semantic similarity, classification, and clustering, with mobile-first retrieval augmented generation pipelines as the headline use case since everything works offline and sensitive data never leaves the device. Weights are published on Hugging Face, Kaggle, and Vertex AI, with inference and finetuning supported through Sentence Transformers. The model ships under Gemma terms that permit responsible commercial use rather than a standard OSI license. For teams that need an open, laptop-and-phone-scale embedder instead of a hosted embedding API, it is Google's flagship option.