이 모델의 강점은 무엇인가요?

Google DeepMind 개발 충분한 8K 컨텍스트 길이 효율적인 벡터 표현 능력

이 모델의 약점은 무엇인가요?

비공개 소스 라이선스 폐쇄적 모델 사양 내부 구조에 대한 외부 접근 없음

어떤 용도에 가장 적합한가요?

시맨틱 검색 구축 문서 유사성 결정 RAG용 벡터 DB 구축

모델 목록으로

Google Deep Mind독점

Gemini Embedding 2

Name: Gemini Embedding 2
Author: Google Deep Mind

Gemini Embedding 2는 Google DeepMind가 개발한 임베딩 모델입니다. 8K의 컨텍스트 길이를 가진 이 모델은 텍스트 표현의 고급 벡터화를 지원합니다.

파라미터

Undisclosed

컨텍스트

라이선스

Proprietary

출시일

2026-03-10

API 가격

이 모델의 API 가격 정보는 현재 공개되지 않았습니다

강점

・Google DeepMind 개발
・충분한 8K 컨텍스트 길이
・효율적인 벡터 표현 능력

약점

・비공개 소스 라이선스
・폐쇄적 모델 사양
・내부 구조에 대한 외부 접근 없음

활용 사례

・시맨틱 검색 구축
・문서 유사성 결정
・RAG용 벡터 DB 구축

심층 분석

Model Type

Multimodal Embedding

Input Token Limit

8,192

Output Dimensions

128-3072 (recommended: 768, 1536, 3072)

Supported Modalities

Text, Image, Video, Audio, PDF

Latest Update

April 2026

Languages

100+

강점

・First natively multimodal embedding model from Google
・State-of-the-art on MTEB Multilingual (69.9) and MTEB Code (84.0)
・Flexible output dimensions via Matryoshka Representation Learning
・Supports 100+ languages for cross-lingual tasks
・Single unified embedding space for all modalities

약점

・Higher cost than text-only embedding models
・Requires Google API access (no open-source weights)
・Large input processing may have latency for video/audio
・Complex multimodal pipelines still need careful orchestration
・Limited fine-tuning options for domain-specific use cases

경쟁사 비교

Model	Arena	SWE	GPQA	Price
Amazon Nova 2 Multimodal	N/A	N/A	N/A	Custom pricing
Voyage Multimodal 3.5	N/A	N/A	N/A	$0.12/1M tokens
OpenAI text-embedding-3-large	N/A	N/A	N/A	$0.13/1M tokens
Cohere Embed v4	N/A	N/A	N/A	$0.10/1M tokens

개요

Gemini Embedding 2 is Google's first natively multimodal embedding model, mapping text, images, video, audio, and PDFs into a single unified embedding space. Released in March 2026 and now generally available, it achieves state-of-the-art performance on cross-modal retrieval benchmarks while supporting 100+ languages.

벤치마크 및 성능

Achieves 69.9 on MTEB Multilingual (vs 68.4 for gemini-embedding-001), 84.0 on MTEB Code (vs 76.0). TextCaps recall@1 of 89.6 for text-to-image, 97.4 for image-to-text. Docci recall@1 of 93.4 for text-to-image. ViDoRe v2 ndcg@10 of 64.9 for text-document retrieval. Significantly outperforms legacy Google models and competitors on most benchmarks.

상세 비교

Outperforms Amazon Nova 2 Multimodal on most benchmarks (e.g., TextCaps 89.6 vs 76.0 text-to-image). Beats Voyage Multimodal 3.5 on cross-modal tasks. Key advantage is native multimodality without separate models per modality. Trade-off is API-only access vs. open-source alternatives.

커뮤니티 평가

Strong positive reception from RAG developers and search engineers. E-commerce and video analysis use cases reported during preview. GA announcement signals production readiness. Developers appreciate the unified embedding space reducing pipeline complexity.

활용 사례

Ideal for multimodal search engines, RAG systems with mixed content, recommendation systems, document retrieval, video analysis, and cross-lingual semantic search. The flexible dimension support (128-3072) allows cost-performance optimization. Particularly valuable for applications needing to search across text, images, and video simultaneously.