Google Releases Multimodal Embedding Model

Google has released "EmbeddingGemma 2," an embedding model that supports not only text but also code, images, videos, and audio. It is an open-weight model distributed under the Apache 2.0 license and is built on the foundation of Gemma 4.

An "embedding model" is a technology that converts data such as text and images into a format (vector) where semantic similarity can be expressed numerically. Through this conversion, it is possible to improve the efficiency of semantic search—quickly finding "data with similar meanings"—and RAG (Retrieval-Augmented Generation), a technique in which AI generates answers while searching external knowledge sources. EmbeddingGemma 2 is designed specifically for this purpose.

Most traditional embedding models have been specialized for text. In contrast, EmbeddingGemma 2's distinguishing feature is its ability to handle multiple information formats (modalities)—code, images, videos, and audio—in a single model. This enables the construction of systems that search and compare different types of data using a single model.

Additionally, this model's performance is optimized for on-device deployment, meaning it is designed to run on local devices rather than in the cloud. The compact design and open-weight nature (where weights are publicly available) represent choices aimed at making it easier for developers to run the model in their own environments.

Publication under the Apache 2.0 license enables utilization across a wide range of applications, including commercial use. Open-weight models offer flexibility distinct from closed APIs in that enterprises and researchers can examine and adjust model internals while using them. This makes them particularly likely to emerge as options for use cases where data should not be sent to the cloud or in privacy-sensitive sectors.

In the context of AI search and RAG, multimodal capability to handle multiple information formats across different types has been a challenge. The emergence of EmbeddingGemma 2 can be viewed as a step forward in lowering barriers to building such systems. Going forward, attention will focus on how use cases for on-device AI expand and how adoption progresses within the developer community.