Google DeepMind Releases EmbeddingGemma 2, a 740M Open Multimodal Embedding Model Built on Gemma 4

Google DeepMind has introduced EmbeddingGemma 2, an open multimodal embedding model built on the Gemma 4 architecture. At 740 million parameters, it is designed to encode five different input modalities into a single 768-dimensional vector space, enabling cross-modal similarity search and retrieval.

Multimodal embeddings are a foundational component for many AI applications, including semantic search, recommendation systems, and retrieval-augmented generation. By unifying multiple input types into one shared space, EmbeddingGemma 2 allows developers to compare and retrieve content across modalities without maintaining separate models for each.

The model's release under Apache 2.0 makes it freely available for both research and commercial use, lowering barriers for teams that want to build multimodal retrieval pipelines without relying on proprietary APIs. Its relatively compact size also makes it practical for deployment on standard infrastructure.

EmbeddingGemma 2 reflects a broader industry trend toward open, lightweight embedding models that compete with offerings from other providers. As multimodal retrieval becomes central to AI product development, accessible models like this could accelerate adoption across startups and enterprise teams alike.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.