Gemini Embedding 2 – Google's first natively multimodal embedding model

Summary: Gemini Embedding 2 maps text, images, video, audio, and documents into a single embedding space, enabling unified multimodal retrieval and classification. It supports multiple media types and languages with flexible input sizes and embedding dimensions, simplifying embedding pipelines.

What it does

It generates embeddings for diverse media including text (up to 8192 tokens), images, video, audio, and PDFs without separate preprocessing. The model supports interleaved inputs and flexible embedding sizes using Matryoshka Representation Learning.

Who it's for

AI developers and ML engineers building search, assistants, knowledge bases, and multimodal AI applications.

Why it matters

It streamlines multimodal AI workflows by replacing fragmented embedding pipelines with a single model that handles multiple media types natively.