EmbeddingGemma 2: Unified and Efficient Multimodal Representations

Henrique Schechter Vera; Sahil Dua; Jovan Andonov; Sara Smoot; Sebastian Russo; Ivan Llanos; Max Matveev; Jae Hyeon Yoo

cs.LGOct 27, 2026

EmbeddingGemma 2: Unified and Efficient Multimodal Representations

Authors: Henrique Schechter Vera, Sahil Dua, Jovan Andonov, Sara Smoot, Sebastian Russo, Ivan Llanos, Max Matveev, Jae Hyeon Yoo, Kat Black, Shih-Cheng Huang, Josiah Aklilu, Maarten Grootendorst, Omar Sanseviero, Chandramouli Amarnath, Ian Ballantyne, Zhe Li, Daniel Salz, Ibrahim Alabdulmohsin, Ilya Philippov, Kristen Quan, Glenn Cameron, Edouard Yvinec, Luciano Martins, Doug Reid, Sindhu Raghuram Panyam, Weiyi Wang, Hengxuan Ying, Madhuri Shanbhogue, Tanmaya Dabral, Shanfeng Zhang, Juyeong Ji, Renjie Wu, Joe Zou, Xu Chen, Yunan Zhang, Yu-Hui Chen, Matthew Soulanille, Vinod Mamillapalli, Yi-Chun Kuo, Wai Hon Law, Matt Kreileder, Alice Zheng, Suril Shah, Mogan Shieh, Pooja Consul, Alberto Montes, Andreas Hess, Andres Sevilla, Babak Samari, Brendan McCloskey, Brian Potetz, Daniel Cer, Georgios Kathareios, Gustavo Hernández Ábrego, Hesen Zhang, Ji Ma, Jiaxing Wu, Keegan Mosley, Paul Suganthan, Rui Meng, Shen Gao, Shengcao Cao, Yi-Ting Chen, Yichang Chen, Zhongli Ding, Kendra Byrne, Cormac Brick, Olivier Lacombe, Qin Yin, Tris Warkentin, Armand Joulin, Tom Duerig, Mojtaba Seyedhosseini

Abstract

We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space. Our approach introduces improvements such as simplified per-layer embeddings and a multi-stage curriculum with modality and objective progression which ensure strong quality across unimodal, cross-modal, and mixed-modal evaluations. Notably, in comprehensive embedding benchmarks across all modalities, EmbeddingGemma 2 establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger. Furthermore, our design offers flexibility for resource-constrained environments, supporting selective modality encoder loading, adjustable vision token budgets, truncatable embedding dimensions, and quantized weights. We present ablations analyzing key decisions and release EmbeddingGemma 2 to promote further research.

Figures & tables

Figure 1: EmbeddingGemma 2 natively maps text, images, video, audio, and combinations thereof into a unified, semantically-aligned vector space.

Figure 2: EmbeddingGemma 2 establishes a new Pareto frontier across benchmarks covering code, image, video, visual documents, and audio. Methods and detailed results in Section 4 . ∗ Mean(Task) for MTEB(Code), MAEB, and MMEB-v2; Mean(Type) for MIEB(lite).

		MTEB(Multilingual, v2)		MTEB(English, v2)		MTEB(Code)	MIEB(lite) ‡		MSEB(Retrieval) ‡
Model	Precision	Mean(Task)	Mean(Type)	Mean(Task)	Mean(Type)	Mean(Task)	Mean(Task)	Mean(Type)	Overall	In-Lang	Cross-Lang
EG 2	bfloat16	61.36	53.82	68.46	63.86	78.68	60.17	65.67	69.54	71.63	67.68
	int4 (Q4_0) ∗	60.96	53.47	68.48	63.91	76.85	56.59	61.95	64.99	68.56	61.80
	int4 (Q4_C) †	60.84	53.36	68.49	63.93	76.70	56.71	62.26	63.92	67.93	60.33
EG 1	bfloat16	61.15	54.31	69.67	65.11	68.76	—	—	—	—	—
EG 1	int4 (Q4_0) ∗	60.62	53.61	69.31	64.65	67.99	—	—	—	—	—

Table 1: Comparison of raw and quantized EmbeddingGemma 2 and EmbeddingGemma 1 checkpoints across benchmarks covering text, vision, and audio: MTEB, MIEB(lite), and MSEB(Retrieval). ∗ Per-block int4. † Per-channel int4. ‡ Different eval harnesses between float and quantized runs.

	MMEB-v2 Overall	Image (Hit@1)	Video (Hit@1)	VisDoc (NDCG@5)
PLE Strategy	MMEB-v2 Overall	Image (Hit@1)	Video (Hit@1)	VisDoc (NDCG@5)
Projection-only	55.46	54.67	47.82	62.37
With lookup	55.07	54.57	47.13	61.76

Table 2: Results with and without a dedicated Per-Layer Embeddings (PLE) lookup table.

	MMEB-v2 Overall	Image (Hit@1)	Video (Hit@1)	VisDoc (NDCG@5)
Training Stage	MMEB-v2 Overall	Image (Hit@1)	Video (Hit@1)	VisDoc (NDCG@5)
Multimodal Pre-finetuning	52.27	50.10	43.26	62.26
$+$ Extra Pre-finetuning	52.28	49.74	43.86	62.40
$+$ Finetuning Stage	55.46	54.67	47.82	62.37

Table 3: Results of continued distillation versus adding a finetuning stage with contrastive learning.

	MMEB-v2 Overall	Image (Hit@1)	Video (Hit@1)	VisDoc (NDCG@5)
Vision Encoder	MMEB-v2 Overall	Image (Hit@1)	Video (Hit@1)	VisDoc (NDCG@5)
Trainable	52.27	50.10	43.26	62.26
Frozen	48.18	44.35	39.24	60.59

Table 4: Results with trainable and frozen vision encoder.

	Vision Token Budget		MMEB-v2 Overall	Image (Hit@1)	Video (Hit@1)	VisDoc (NDCG@5)
Experiment	Train (Img / Vid)	Eval	MMEB-v2 Overall	Image (Hit@1)	Video (Hit@1)	VisDoc (NDCG@5)
Inference-Time Scaling	280 / 140	0 70	50.07	50.90	46.33	51.61
	280 / 140	140	55.37	53.58	47.99	63.60
	280 / 140	280	56.15	55.15	48.51	63.38
	280 / 140	560	57.02	55.21	46.52	67.62
Train-Time Scaling	280 / 0 70	0 70	49.62	50.56	46.60	50.49

Table 5: Results with different vision token budgets (max soft tokens per image or video frame) at train and inference time. Evaluation uses a single budget for all tasks. Top: Inference-time scaling of a model trained with 280 tokens per image and 140 per video frame. Bottom: Train-time scaling with 70 vs. 140 tokens per video frame. The two experiments use different intermediate recipes.

		Embedding Gemma 2	Embedding Gemma 1	Jina Embed v5 Omni Nano	Qwen3 Embed 0.6B	InterestFM UIR CAFe 0.5B	SigLIP SoViT 400M	LAION CLAP General
Total Parameters		744M	308M	986M	596M	894M	878M	194M
Embedding Dimension		768	768	768	1024	896	1152	512
MTEB(Multi, v2) [ 55 ]	Mean (Task)	61.36	61.15	65.52	64.34	–	–	–
MTEB(Multi, v2) [ 55 ]	Mean (Type)	53.82	54.31	57.66	56.01	–	–	–
MTEB(Eng, v2) [ 55 ]	Mean (Task)	68.46	69.67	71.11	70.47	–	–	–
MTEB(Eng, v2) [ 55 ]	Mean (Type)	63.86	65.11	65.22	64.72	–	–	–

Table 6: Comparison on benchmarks against a slightly larger multimodal peer and modality specialists.

*

Modality not supported. – Score unavailable. † Computed by transcribing audio with ASR.

Model Name	Params	Dim.	Mean (Task)	AppsR.	COIR	CESR	CFMT	CSMT	CSNCCR	CSNR	CTOC	CTODL	CQA	SOQA	ST2SQL
EmbeddingGemma 2	271M	768	78.7	89.2	86.9	80.5	85.6	82.2	93.8	90.9	84.6	35.6	53.3	89.1	72.5
		512	77.2	89.2	86.6	79.9	85.3	82.0	93.7	90.7	82.9	35.7	52.7	87.0	61.1
		256	76.2	88.3	85.5	77.9	84.1	80.5	92.9	90.2	82.8	35.8	50.5	84.9	60.9
		128	71.4	82.5	80.8	70.0	78.5	74.8	89.5	88.1	77.4	34.6	44.1	77.0	59.7
pplx-embed-v1-0.6b	596M	1024	75.8	78.7	82.4	76.4	83.0	84.7	88.1	89.9	85.0	35.2	43.3	90.7	73.0
UltraBERT-Embed-150M	151M	768	73.3	68.5	82.2	62.2	84.4	86.4	93.7	90.0	88.1	37.2	36.3	92.4	58.0

Table 7: Performance of top-ranking models under 1B parameters on MTEB(Code).

			Mean	Image					Video					VisDoc
Model Name	Params	Dim.	(Task)	CLS	QA	RET	GD	All	CLS	QA	RET	MRET	All	VDRv1	VDRv2	VR	OOD	All
EmbeddingGemma 2	439M	768	59.0	54.4	41.9	65.8	77.4	57.3	56.2	48.6	47.2	50.6	50.7	70.6	58.3	72.7	63.3	67.8
		512	58.4	54.2	41.4	65.4	77.1	56.9	56.0	48.5	46.5	50.3	50.3	69.2	56.3	72.0	62.2	66.6
		256	56.2	53.2	39.5	63.6	76.5	55.5	54.6	48.4	44.9	49.2	49.3	64.9	53.2	67.2	59.5	62.6
		128	45.6	39.9	29.6	54.7	72.8	45.6	40.8	44.4	34.9	43.6	40.6	51.9	39.2	52.6	48.8	49.5
Eddy-vl_embed_1.9B_v1	1.93B	2048	63.2	54.9	61.7	64.8	84.7	63.4	51.5	57.9	41.9	48.0	50.0	78.8	55.7	79.1	64.9	72.7

Table 8: Performance of top-ranking models under 2B parameters on MMEB-v2.

			Mean (Task)	Mean (Type)	Multi. Retr.	A2A Retr.		Doc. Und.	Image Class.	Image Clus.	Vision QA	VSTS (eng)	VSTS (multi)	0-shot Class.
Model Name	Params	Dim.	Mean (Task)	Mean (Type)	Multi. Retr.	A2A Retr.	Comp.	Doc. Und.	Image Class.	Image Clus.	Vision QA	VSTS (eng)	VSTS (multi)	0-shot Class.
EmbeddingGemma 2	439M	768	60.2	65.7	65.4	38.3	45.7	67.3	74.3	81.3	59.1	81.5	76.8	67.0
		512	59.9	65.3	64.8	38.1	45.7	66.5	74.2	80.0	58.8	81.6	76.8	66.9
		256	58.6	64.2	62.9	37.1	44.5	64.0	73.7	78.9	56.4	81.8	76.8	65.5
		128	54.2	60.0	50.7	33.0	40.1	56.6	71.5	79.3	56.7	80.8	74.8	56.4
siglip-so400m-patch14-384	878M	1152	52.3	53.5	35.4	32.4	32.8	46.9	78.8	75.9	48.0	69.6	41.4	73.8

Table 9: Performance of top-ranking models under 1B parameters on MIEB(lite).

			Mean (Task)	Mean (Type)	A2A Retr.	Audio Class.	Audio Clus.	Multi. Class.	Pair Class.	Audio Rerank.	0-shot Class.
Model Name	Params	Dim.	Mean (Task)	Mean (Type)	A2A Retr.	Audio Class.	Audio Clus.	Multi. Class.	Pair Class.	Audio Rerank.	0-shot Class.
EmbeddingGemma 2	577M	768	49.9	48.7	53.6	52.9	4.1	38.7	67.4	68.5	55.4
		512	49.8	48.6	53.4	52.7	4.3	38.8	67.4	68.8	55.0
		256	49.1	48.2	52.1	52.2	4.2	38.4	67.2	68.7	54.8
		128	47.3	46.3	49.0	50.4	4.5	30.8	66.9	67.1	55.1
larger_clap_general	194M	512	35.0	41.1	18.9	47.9	6.6	80.6	51.9	66.8	14.9

Table 10: Performance of top-ranking models under 1B parameters on MAEB.

Model Name	Audio Mode	Parameters	Dim.	Overall	In-Lang	Cross-Lang
EmbeddingGemma 2	Native Audio	577M	768	69.54	71.63	67.68
			512	69.18	71.20	67.38
			256	66.76	68.86	64.89
			128	56.71	60.30	53.49
EmbeddingGemma 2	ASR	271M	768	57.59	62.14	53.52
EmbeddingGemma 1	ASR	308M	768	50.95	61.10	41.87

Table 11: Results with ASR and native audio on MSEB(Retrieval) (MRR@10).

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Use Case	Query Task Instruction	Target Task Instruction
Web / Document Search	task: search result \| query: {text}	title: {title} \| text: {text}
Question Answering	task: question answering \| query: {text}	title: {title} \| text: {text}
Fact Checking	task: fact checking \| query: {text}	title: {title} \| text: {text}
Code Search	task: code retrieval \| query: {text}	title: {title} \| text: {text}
Classification	task: classification \| query: {text}	Same as query
Clustering	task: clustering \| query: {text}	Same as query

Appendix

Table 12: Task instruction prefixes used for EmbeddingGemma 2. Asymmetric tasks (top) format queries and targets differently, using title: none when no document title is available. Symmetric tasks (bottom) apply the same prefix to all inputs.

			Mean (Task)	Mean (Type)	Bitext Mining			Inst. Retrieval	Multi. Class.	Pair Class.
Model Name	Parameters	Dim.	Mean (Task)	Mean (Type)	Bitext Mining	Class.	Clus.	Inst. Retrieval	Multi. Class.	Pair Class.	Rerank.	Retrieval	STS
EmbeddingGemma 2	271M	768	61.4	53.8	65.7	63.0	50.4	2.9	25.2	78.8	62.9	60.8	74.7
		512	61.2	53.7	65.4	62.7	51.0	2.5	25.1	78.8	62.8	60.2	74.4
		256	60.4	53.0	63.6	61.9	51.1	2.0	24.7	78.7	62.0	58.7	74.1
		128	57.9	50.7	59.4	59.7	49.0	0.8	23.3	78.5	59.9	52.3	73.6
harrier-oss-v1-0.6b	596M	1024	69.0	59.0	82.8	73.9	54.0	0.8	26.4	82.1	63.2	70.8	77.1

Appendix

Table 13: Performance of top-ranking models under 1B parameters on MTEB(Multilingual, v2).

			Mean (Task)	Mean (Type)			Pair Class.
Model Name	Parameters	Dim.	Mean (Task)	Mean (Type)	Class.	Clus.	Pair Class.	Rerank.	Retrieval	STS	Summ.
EmbeddingGemma 2	271M	768	68.5	63.9	88.8	55.4	86.4	46.3	51.7	83.2	35.1
		512	68.4	63.9	88.7	55.3	86.4	46.3	51.6	83.3	35.6
		256	67.8	63.3	88.5	55.5	86.1	45.9	49.4	83.1	34.8
		128	65.7	61.5	87.6	54.3	85.3	45.7	43.7	82.5	31.4
jina-embed-v5-text-nano	212M	768	71.1	65.2	90.1	53.5	84.7	49.2	58.8	88.3	31.9

Appendix

Table 14: Performance of top-ranking models under 1B parameters on MTEB(English, v2).

Task Name	Performance
AILAStatutes	32.44
AfriSentiClassification	45.67
AlloProfClusteringS2S.v2	52.31
AlloprofReranking	76.92
AmazonCounterfactualClassification	85.83
ArXivHierarchicalClusteringP2P	63.42

Appendix

Table 15: Results of EmbeddingGemma 2 on MTEB(Multilingual, v2).

Task Name	Performance
AmazonCounterfactualClassification	91.10
ArXivHierarchicalClusteringP2P	63.66
ArXivHierarchicalClusteringS2S	60.59
ArguAna	59.95
AskUbuntuDupQuestions	60.61
BIOSSES	85.33

Appendix

Table 16: Results of EmbeddingGemma 2 on MTEB(English, v2) (left) and MTEB(Code) (right).

Task Name	Performance
BeijingOpera	81.37
BirdCLEF	16.00
CREMADPairClassification	52.50
CREMA_D	19.67
CREMA_DClustering	0.57
ClothoT2ARetrieval	10.71

Appendix

Table 17: Results of EmbeddingGemma 2 on MAEB (left) and MSEB(Retrieval) (right).

Task Name	Performance
AROCocoOrder	57.00
AROFlickrOrder	56.12
AROVisualAttribution	77.97
AROVisualRelation	56.93
BLINKIT2IMultiChoice	75.87
CIFAR100ZeroShot	75.61

Appendix

Table 18: Results of EmbeddingGemma 2 on MIEB(lite).

Task Name	Performance
A-OKVQA	35.20
ActivityNetQA	52.10
Breakfast	35.57
CIRR	40.10
Charades-STA	25.58
ChartQA	18.90

Appendix

Table 19: Results of EmbeddingGemma 2 on MMEB-v2.

Multimodal Embedding

Multimodal embedding maps text, images, audio, video, and other data into representations that support comparison and retrieval across modalities.

See all 160 papers

Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini
May 26, 2026Madhuri Shanbhogue, Zhe Li, Shanfeng Zhang +86Cross-Modal Learning Unified Multimodal Models
jina-embeddings-v5-omni: Geometry-preserving Embeddings via Locked Aligned Towers
May 8, 2026Florian Hönicke, Michael Günther, Andreas Koukounas +4Cross-Modal Learning Cross-Modal Representation Learning
Jina by Elastic
Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings
Sep 21, 2026Ovis-Embedding TeamUnified Multimodal Models Multimodal Embedding
Alibaba Token Hub, Alibaba Group

May 26, 2026J/K move · Enter open · S save

Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini

Madhuri Shanbhogue, Zhe Li, Shanfeng Zhang, Gustavo Hernández Ábrego+85

Cross-Modal Learning Unified Multimodal Models Multimodal Embedding Cross-Modal Retrieval Multimodal Foundation Models Multimodal Understanding

Open paper AskPDF arXiv

We introduce Gemini Embedding 2, a native multimodal embedding model that allows embedding video, audio, image, and text modalities in a unified representation space. We leverage the multimodal capabilities of Gemini to produce embeddings for arbitrary combinations of interleaved inputs across all these modalities that generalize well across a wide variety of tasks. Applying large-scale contrastive learning in a multi-task multi-stage training setup, we achieve state-of-the-art performance on key embedding benchmarks including unimodal, cross-modal, and multimodal retrieval spanning a diverse set of tasks. We show that our embedding model demonstrates strong performance (with a score of 62.9 R@1 on MSCOCO, 68.8 NDCG@10 on Vatex, 69.9 on MTEB multilingual and 84.0 on MTEB Code) across a variety of tasks surpassing the performance of specialized models. These unified capabilities make Gemini Embedding 2 a promising candidate for downstream use cases such as RAG, recommendation and search. Furthermore, its robust zero-shot performance across distinct fields - from astronomy and bioscience to fine arts and the culinary arts - establishes it as a highly reliable, out-of-the-box representation even for specialized domains.

Abstract

Multimodal Embedding

Explore similar work

Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini

jina-embeddings-v5-omni: Geometry-preserving Embeddings via Locked Aligned Towers

Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings