Gemini Embedding 2: First Natively Multimodal Embedding Model

Google has released Gemini Embedding 2, its first natively multimodal embedding model that maps text, images, videos, audio, and documents into a single unified embedding space across 100+ languages. The model supports flexible output dimensions via Matryoshka Representation Learning (scaling down from 3072), handles interleaved multi-modal input in a single request, and processes up to 8192 text tokens, 6 images, 120 seconds of video, native audio, and 6-page PDFs. It establishes state-of-the-art performance across text, image, video, and speech embedding benchmarks, outperforming leading models. Early access partners report significant improvements including 85.3% text-to-video Recall@1 at Paramount Skydance and up to 70% latency reduction at Sparkonomy by eliminating intermediate LLM inference steps.

Source: Gemini Embedding 2: Our first natively multimodal embedding model