Google has launched EmbeddingGemma 2, an open multimodal embedding model designed to operate locally on smartphones. Unlike its predecessor, which was limited to text, the new version integrates images, audio, and video into a single embedding space.
The model is built on the Gemma 4 architecture and contains 740 million parameters, more than double the size of the original version. The text-only core consists of 270 million parameters and requires approximately 191 megabytes of memory when quantized.
Developers can utilize Matryoshka Representation Learning to compress embeddings from 768 numbers down to 128, reducing storage requirements by up to six times. According to Google, maintaining a length of 256 preserves about 95% of the original quality for retrieval tasks.
In benchmarks, EmbeddingGemma 2 achieved a score of 78.68 on the code section of the Massive Text Embedding Benchmark, marking an improvement of nearly 10 points over the first version. Google reports that the model outperforms several larger specialist models.
The model is available under an Apache 2.0 license via Hugging Face and Kaggle, allowing for commercial use. It is compatible with tools like vLLM, llama.cpp, and Ollama, and will soon be added to the Model Garden in the Gemini Enterprise Agent Platform.
