Gemini Embedding 2

Embedding modelCard summarized

Our first natively multimodal embedding model

Source: Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini (technical report, arXiv:2605.27295)

Use it via API

$0.20 / 1M tok

Code example Compare

Where to call it

Gemini APIgemini-embedding-2
Google Vertex AIgemini-embedding-2

Supports

Batch API
Image input
PDF input

Full feature support

Good for

Cross-modal semantic search (text to image/video/audio and back)
Document retrieval, including PDFs with OCR
Recommendations

Overview

Google · 02Embedding

Gemini Embed 2

Our first natively multimodal embedding model

Typical price

$0.20 / 1M tok

Text input

Dimensions: 3072
Max input: 8,192 tokens shared across modalities
Modes: text embeddings, image embeddings
Released: 22 Apr 2026

Gemini Embedding 2 is Google's natively multimodal embedding model. It is initialized from Gemini and maps text, images, video, audio and documents into one shared vector space of up to 3,072 dimensions. Google reports state-of-the-art results on unimodal, cross-modal and multimodal retrieval, and gains over its previous text-only Gemini Embedding on multilingual and code benchmarks. It launched in preview on 10 Mar 2026 and became generally available on 22 Apr 2026. technical report abstract, §2; Gemini API changelog

gemini-embedding-2

Typical price: $0.20 / 1M tok
Dimensions: 3072 (can be shortened to 128–3072; 768, 1536, 3072 recommended)
Max input: 8,192 tokens shared across modalities (up to 6 images, 120 s video, 180 s audio, 1 PDF of 6 pages)
Languages: 100+ languages
Weights: Closed
Released: 22 Apr 2026

text embeddings
image embeddings
audio embeddings
video embeddings
PDF embeddings
multimodal embeddings (interleaved text + media)
code retrieval