Google DeepMind Releases EmbeddingGemma 2, a 740 Million Parameter Open-Weight Multimodal Model for On-Device Search

Google DeepMind launched EmbeddingGemma 2, a 740 million parameter multimodal AI model designed to enable on-device search and retrieval without relying on cloud servers. The model is released as open-weight software under the Apache 2.0 license, allowing developers to freely use, modify, and distribute it in commercial and private applications.

EmbeddingGemma 2 is built to handle multimodal inputs, processing text and other data types to power search and retrieval tasks directly on personal devices. By running locally, it preserves user privacy and reduces dependence on server infrastructure, which the company and analysts say could open significant new possibilities for privacy-focused AI applications on smartphones and other edge hardware.

The Apache 2.0 licensing terms are notable because they impose minimal restrictions on commercial use, lowering the barrier for developers and companies to integrate the model into products. The release adds to Google's broader Gemma family of open-weight models and positions Alphabet, ticker GOOGL, to compete more directly with other open-weight AI providers in the on-device segment.