Google released EmbeddingGemma 2 on October 6, 2026, saying it converts text, images, video, audio, and code into numerical vectors for easier content comparison. It is the company's first major update to its embedding models since the launch of EmbeddingGemma 1 in 2024.
Google reported a benchmark score of 78.68 on the Massive Text Embedding Benchmark (Code), measured on a multimodal dataset. That compares with a score of 68.76 from its predecessor, a jump of nearly 10 points.
EmbeddingGemma 2 is built on WebGPU technology and targets local, offline applications requiring minimal computational resources. Availability begins immediately, initially for developers and researchers.
"Each query takes about 20 to 70 milliseconds via WebGPU in the browser," said Matthias Bastian, a contributor at thedecoder. "The model needs only around 191 MB of RAM and cuts local vector database storage by up to six times."
The announcement follows Google's recent focus on open-source AI tools and local model deployment. Google itself frames the significance as a step toward more accessible and efficient embedding solutions.
Google did not say how the model will perform on non-text tasks, and it raises the question of whether the gains seen on text benchmarks will translate to other modalities. The weights are available on Hugging Face and Kaggle, along with a developer guide and documentation.
Source: thedecoder
