Google today released EmbeddingGemma, a 308-million-parameter open-weights text embedding model built on the Gemma 3 architecture. The model is designed to run directly on edge devices such as mobile phones, laptops, and desktops, using less than 200MB of RAM with quantization. It supports over 100 languages and delivers customizable output dimensions from 768 down to 128 via Matryoshka representation, with a 2K token context window.

Google claims EmbeddingGemma is the highest-ranking open multilingual text embedding model under 500 million parameters on the MTEB benchmark, with inference under 15 milliseconds for 256 tokens on EdgeTPU. The model is engineered for privacy-preserving on-device applications including retrieval-augmented generation and semantic search, operating entirely offline. Weights are available on Hugging Face, Kaggle, and Vertex AI, with integrations already supported across tools such as sentence-transformers, llama.cpp, MLX, Ollama, LiteRT, transformers.js, and LangChain.

“EmbeddingGemma is comparable to popular models nearly twice its size,” the team wrote. The release comes as Google pushes mobile-first AI pipelines, pairing EmbeddingGemma with its on-device Gemma 3n generative model so that both retrieval and generation can run locally, without a network round-trip.