Summary
Google DeepMind has introduced EmbeddingGemma, a 308-million-parameter open embedding model designed to run on everyday devices. The model ranks first among open multilingual embedding models under 500M parameters on the MTEB benchmark and supports offline RAG and semantic search.
On September 4, 2025, Google DeepMind released EmbeddingGemma, an open embedding model designed for on-device artificial intelligence. The model converts text into numerical representations—embeddings—that capture meaning, enabling tasks such as semantic search and retrieval-augmented generation (RAG) without an internet connection.
Contents
What is EmbeddingGemma?
EmbeddingGemma is a text embedding model built on the Gemma 3 architecture. It has 308 million parameters, which the company splits into roughly 100 million model parameters and 200 million embedding parameters. Trained on more than 100 languages, it is intended to produce high-quality multilingual embeddings.
The model supports a context window of up to 2,000 tokens, allowing it to process short documents or conversations in a single pass. Its output dimensionality is flexible: using Matryoshka Representation Learning, developers can extract vectors of 768, 512, 256 or 128 dimensions from the same model, trading off between accuracy and storage cost.
Performance and efficiency
Google states that EmbeddingGemma is the highest-ranking open multilingual text embedding model under 500 million parameters on the Massive Text Embedding Benchmark (MTEB). The company also claims that it performs comparably to popular embedding models with nearly twice as many parameters.
For on-device use, the model is engineered to run in less than 200 MB of RAM when quantized. On an Edge TPU, the reported inference time is below 15 milliseconds for a 256-token input, which the company says enables real-time interactions. The combination of quantization-aware training and the compact architecture is intended to make the model suitable for mobile phones, laptops and desktops.
Offline and privacy
A central design goal of EmbeddingGemma is offline operation. Because embeddings are generated directly on the user's hardware, no data needs to leave the device. Google highlights use cases such as searching personal files, texts, emails and notifications without an internet connection, and building offline chatbots through RAG with Gemma 3n. The model can also be fine-tuned for specific domains or languages using a provided quickstart notebook.
Integration and availability
EmbeddingGemma is compatible with many existing tools, including sentence-transformers, llama.cpp, MLX, Ollama, LiteRT, transformers.js, LMStudio, Weaviate, Cloudflare, LlamaIndex and LangChain. Model weights are available on Hugging Face, Kaggle and Vertex AI. Documentation, inference and fine-tuning guides, and a RAG example are published alongside the model. A technical report describing the research is also available.