Gemini Embedding 2 Preview

• 🆕 Google's natively multimodal embedding model, offered in public preview.
• 🌐 Maps text, images, video, audio, and documents into one space.
• 📏 Generates output vectors up to 3,072 dimensions.
• 🔧 Flexible output sizes via Matryoshka representation learning.
• 🎯 Built for cross-modal semantic search, retrieval, and RAG.
• 🏢 Available through the Gemini API and Vertex AI.
• 🧠 Inherits multimodal understanding from the Gemini foundation model.

Google

Google is an American multinational technology corporation and one of the world's most valuable brands. A subsidiary of parent company Alphabet Inc., Google operates across search, cloud computing, consumer electronics, and artificial intelligence. Its…

Gemini Embedding 2 Preview is Google's embedding model released in 2026, documented as a natively multimodal embedding model. Rather than handling a single modality, it projects text, images, video, audio, and documents into one shared vector space, enabling cross-modal retrieval, classification, and recommendation. It is offered through the Gemini API and Vertex AI and, at the time of writing, is in public preview.

The most concrete generational change is scope. Google's earlier embedding work centered on text, whereas Gemini Embedding 2 inherits multimodal understanding from the Gemini foundation model, extending the family beyond text-only retrieval. It produces vectors up to 3,072 dimensions for downstream search and similarity tasks.

A practical feature is Matryoshka Representation Learning, which concentrates the most important semantic information into the leading dimensions, so developers can truncate to smaller vector sizes to save storage with limited quality loss. This lets teams tune the trade-off between index size and retrieval quality without retraining.

Within the wider Gemini ecosystem, it sits alongside generative siblings like Gemini 3.5 Flash and Gemini 3.1 Pro Preview, serving as the retrieval layer for RAG pipelines that those text models can consume.