jina-embeddings-v5-text: Task-Targeted Embedding Distillation

Mohammad Kalim Akram; Saba Sturua; Nastia Havriushenko; Quentin Herreros; Michael Günther; Maximilian Werk; Han Xiao

cs.CLFeb 17, 2026

jina-embeddings-v5-text: Task-Targeted Embedding Distillation

Authors: Mohammad Kalim Akram, Saba Sturua, Nastia Havriushenko, Quentin Herreros, Michael Günther, Maximilian Werk, Han Xiao

Organizations: Jina by Elastic

Abstract

Text embedding models are widely used for semantic similarity tasks, including information retrieval, clustering, and classification. General-purpose models are typically trained with single- or multi-stage processes using contrastive loss functions. We introduce a novel training regimen that combines model distillation techniques with task-specific contrastive loss to produce compact, high-performance embedding models. Our findings suggest that this approach is more effective for training small models than purely contrastive or distillation-based training paradigms alone. Benchmark scores for the resulting models, jina-embeddings-v5-text-small and jina-embeddings-v5-text-nano, exceed or match the state-of-the-art for models of similar size. jina-embeddings-v5-text models additionally support long texts (up to 32k tokens) in many languages, and generate embeddings that remain robust under truncation and binary quantization. Model weights are publicly available, hopefully inspiring further advances in embedding model development.

Figures & tables

Figure 1: Architecture of jina-embeddings-v5-text .

Model	Parameters		RoPE	Max.	Emb.
Name	Base	LoRA	$\bm{\theta}$	Tokens	Dim.
j-v5-text-small	596M	4 $\times{}$ 20.2M	3.5M	32K	1024
j-v5-text-nano	212M	4 $\times{}$ 6.7M	1M	8K	768
Base Models
Qwen3-0.6B	600M	–	1M	32K	1024
EuroBERT-210M	210M	–	250K	8K	768

Table 1: Attributes of the Base Models and the Resulting Embedding Models

Model	Params	Dim	Avg Tasks	Avg Type	BM	Cls	Clu	IR	MLC	Pair	RR	Ret	STS
Qwen3-4B	4B	2560	69.5	60.9	79.4	72.3	57.1	11.6	26.8	85.1	65.1	69.6	80.9
jina-v4	3.8B	2048	58.17	51.55	62.4 †	55.2 †	44.0 †	0.7 †	19.3 †	79.3 †	62.20 †	66.4	74.4
Qwen3-0.6B (instr.)	596M	1024	64.3	56.0	72.2	66.8	52.3	5.1	24.6	80.8	61.4	64.7	76.2
Qwen3-0.6B (generic)	596M	1024	61.1	54.3	72.2	58.4 †	49.8 †	3.8 †	21.1 †	80.8	62.2 †	64.2 †	76.2
jina-v3	572M	1024	58.4	50.7	65.3	58.8	45.6	-1.3	18.4	79.3	57.1	55.8	77.1
snowflake-l-v2	568M	1024	57.0	50.0	64.1	57.4	42.8	-2.5	18.9	76.7	63.7	58.4	70.1

Table 2: MTEB (Multilingual, v2) Evaluation Results

Figure 2: Performance of j-v5-text-small on different languages on MMTEB compared to other models

Model	Params	Dim	Avg Tasks	Avg Type	Cls	Clu	Pair	RR	Ret	STS	Sum
Qwen3-4B	4B	2560	74.6	68.1	89.8	57.5	87.0	50.8	68.5	88.7	34.4
jina-v4	3.8B	2048	65.09	60.68	74.1 †	45.5 †	83.1 †	48.04 †	56.2	85.9	32.0 †
Qwen3-0.6B (instr.)	596M	1024	70.5	64.7	84.6	54.1	84.4	48.2	61.8	86.6	33.4
Qwen3-0.6B (generic)	596M	1024	67.0	62.0	72.0 †	51.8 †	84.4	46.2 †	59.8 †	86.6	33.4
jina-v3	572M	1024	65.7	62.6	85.8	47.4	84.0	47.9	54.3	85.8	32.9
snowflake-l-v2	568M	1024	63.6	59.8	73.4	44.4	83.0	47.5	58.6	78.1	33.8

Table 3: MTEB(eng, v2) Evaluation Results

Model	Params	Dim	Avg Tasks	MTEB-M	MTEB-E	RTEB	BEIR	Long
Qwen3-4B	4B	2560	67.95	69.60	68.46	70.77 †	61.58	78.82 †
jina-v4	3.8B	2048	63.62	66.43	56.15	66.52	53.97 †	69.88
Qwen3-0.6B	596M	1024	61.87	64.65	61.83	64.21 †	55.52	72.20 †
jina-v3	572M	1024	56.11	55.76	54.29	54.58 †	53.17	55.67
snowflake-l-v2	568M	1024	57.59	58.36	58.56	53.95	55.22	63.74
mult.-e5-l-instr.	560M	1024	54.22	57.12	53.47	54.78	52.74	41.76

Table 4: Retrieval Benchmark Results

Figure 3: Performance comparison of different training objectives. Average nDCG@10 on the MTEB (English, v2) benchmark for S2ORC (left) and the full training data mixture (right).

Figure 4: Comparison of projection configurations on S2ORC . Performance is measured by average nDCG@10 on MTEB (English, v2).

Loss Configuration	MTEB	RTEB
$\mathcal{L}_{\text{NCE}}+\mathcal{L}_{\text{distill}}+\mathcal{L}_{\text{GOR}}$	64.50	66.45
$\mathcal{L}_{\text{NCE}}+\mathcal{L}_{\text{distill}}$	64.21	66.16
$\mathcal{L}_{\text{NCE}}+\mathcal{L}_{\text{GOR}}$	64.11	66.11
$\mathcal{L}_{\text{distill}}+\mathcal{L}_{\text{GOR}}$	63.49	65.05
$\mathcal{L}_{\text{NCE}}$	63.38	65.14
$\mathcal{L}_{\text{distill}}$	63.16	64.37

Table 5: Evaluation of retrieval adapter training losses on MTEB v2 Retrieval subset and public RTEB tasks.

	MTEB		RTEB
Configuration	BF16	Binary	BF16	Binary
Full (w/ GOR)	64.50	62.60 (-1.90)	66.45	63.94 (-2.51)
w/o GOR	64.21	61.13 (-3.08)	66.16	62.24 (-3.92)

Table 6: Impact of GOR loss on quantization robustness, evaluated on MTEB v2 Retrieval subset and public RTEB tasks.

Figure 5: Average MMTEB score across reduced embedding dimensions.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Model	Steps	Devices & Batch size	Max. Tokens (Seq. Length)	LR	$\theta$	Others
First Stage	j-v5-text-small	50,000	8 $\times{}$ 512	512	$1\cdot{}10^{-4}$	1M
First Stage	j-v5-text-nano	50,000	8 $\times{}$ 1024	512	$1\cdot{}10^{-4}$	250K
Long Context	j-v5-text-small	6,500	2 $\times{}$ 64	4096	$1\cdot{}10^{-4}$	500K
Training	j-v5-text-small	6,500	2 $\times{}$ 64	4096	$1\cdot{}10^{-4}$	500K
Asym. Retrieval	j-v5-text-small	8,000	2 $\times{}$ (256 / 64)	384 / 4096	$2\cdot{}10^{-5}$	1M	$\tau=0.02,\lambda_{D}=2,$
Asym. Retrieval	j-v5-text-nano	8,000	2 $\times{}$ (384 / 96)	384 / 4096	$2\cdot{}10^{-5}$	250K	$\lambda_{\text{NCE}}=\lambda_{S}=1$

Appendix

Table A1: Hyperparameters for the different models and training stages.

Model	AVG	Arg	CQG	CQU	CFHN	FEV	FiQA	HPQA	SCI	TREC	TOU
Qwen3-4B	68.46	75.64	71.51	59.60	48.48	92.47	62.65	75.22	31.44	92.92	74.65
jina-v4	56.15	66.96	57.59	42.95	34.57	87.69	48.49	69.01	21.48	80.36	52.41
Qwen3-0.6B (instr.)	61.83	70.97	64.14	51.49	43.62	88.94	46.61	67.69	24.41	90.52	69.90
Qwen3-0.6B (generic)	59.77	67.48	63.83	49.80	37.61	86.66	46.83	66.90	22.77	88.80	67.01
jina-v3	54.29	43.29	58.02	43.52	43.14	89.90	47.35	64.70	19.92	77.74	55.28
snowflake-l-v2	58.56	59.11	63.18	46.57	42.83	92.21	45.35	68.40	20.28	83.63	64.05

Appendix

Table A2: Evaluation Results on MTEB Retrieval Tasks (nDCG@10 [%])

Model	Avg	MindSmallReranking	AskUbuntuDupQuestions
Qwen3-4B	50.76	32.71	68.81
jina-v4	48.04	32.51	63.56
Qwen3-0.6B (instr.)	48.18	31.23	65.13
Qwen3-0.6B (generic)	46.17	31.40	60.94
jina-v3	47.94	30.83	65.04
snowflake-l-v2	47.47	31.59	63.35

Appendix

Table A3: Evaluation Results on MTEB Reranking Tasks (MAP@1000 [%]).

Model	Avg	BIO	SICK-R	STS12	STS13	STS14	STS15	STS17	STS22	STSB
Qwen3-4B	88.7	82.9	88.1	86.6	94.4	90.9	93.8	95.1	73.1	93.7
jina-v4	85.9	89.2	89.2	83.5	88.6	84.8	89.7	88.7	70.7	88.6
Qwen3-0.6B (instr.)	86.6	85.5	84.8	83.0	91.8	87.1	91.4	93.3	71.1	91.1
Qwen3-0.6B (generic)	86.6	85.5	84.8	83.0	91.8	87.1	91.5	93.2	71.1	91.1
jina-v3	85.8	88.7	89.6	82.4	89.5	85.0	89.3	90.0	68.4	89.4
snowflake-l-v2	78.0	87.2	74.0	71.2	80.4	75.1	82.9	84.5	67.9	79.2

Appendix

Table A4: Evaluation Results on MTEB Semantic Textual Similarity Tasks (Spearman correlation [%]).

Model	Avg	SprintDuplicateQuestions	TwitterSemEval2015	TwitterURLCorpus
Qwen3-4B	87.0	96.1	77.8	87.2
jina-v4	83.1	91.4	71.5	86.4
Qwen3-0.6B (instr.)	84.4	94.1	72.3	86.8
Qwen3-0.6B (generic)	84.4	94.1	72.3	86.8
jina-v3	84.0	97.0	70.9	84.1
snowflake-l-v2	83.0	96.5	67.0	85.4

Appendix

Table A5: Evaluation Results on MTEB Pair Classification Tasks (Max Average Precision [%]).

Model	Avg	AmzCF	Bnk77	IMDB	MTOP	M-Int	M-Scn	Tox	TwSent
Qwen3-4B	89.8	93.7	86.3	97.2	97.8	85.0	88.8	91.4	78.4
jina-v4	74.1	72	78.6	81.0	93.2	70.5	72.6	64.5	60.3
Qwen3-0.6B (instr.)	84.6	91.5	81.0	95.4	96.0	80.4	74.1	82.1	76.0
Qwen3-0.6B (generic)	72.0	70.1	71.6	89.6	91.0	50.0	74.4	68.0	61.5
jina-v3	85.8	90.9	84.1	91.9	97.5	75.2	84.1	91.3	71.4
snowflake-l-v2	73.4	65.6	81.8	72.8	93.5	71.5	76.2	65.9	59.6

Appendix

Table A6: Evaluation Results on MTEB Classification Tasks (Accuracy [%]).

Model	Avg	AXP	AXS	BioP	MedP	MedS	SEx	SExP	20News
Qwen3-4B	57.5	64.8	65.2	50.9	45.3	43.8	77.6	53.5	59.0
jina-v4	45.5	58.2	57.3	38.4	35.0	34.8	53.6	39.8	47.0
Qwen3-0.6B (instr.)	54.1	63.7	63.8	47.3	42.2	40.4	71.2	52.1	51.7
Qwen3-0.6B (generic)	51.8	63.2	61.8	45.1	40.7	39.8	66.2	45.1	52.4
jina-v3	47.4	58.9	55.9	40.0	38.2	37.2	56.7	40.9	51.5
snowflake-l-v2	44.4	57.2	53.1	37.2	35.0	32.6	56.3	41.2	42.4

Appendix

Table A7: Evaluation Results on MTEB Clustering Tasks (V-measure [%] - see footnote in Section 5.1 ).

Model	Avg	AI	Arg	Bel	Cov	Hag	PK	LB	MIR	ML	SD	SQA	SO	STC	TC	TR	TW	Wiki	WG
Qwen3-4B	69.6	81.2	75.6	81.2	87.4	98.8	84.3	95.4	69.5	81.9	31.4	20.2	94.3	42.3	92.9	1.2	72.6	91.2	51.5
jina-v4	66.4	50.1	67.0	74.3	80.2	98.8	69.8	94.8	62.9	74.9	21.5	30.2	91.4	58.1	80.4	1.3	84.4	88.5	67.3
Qwen3-0.6B (instr.)	64.6	79.0	71.0	68.7	84.8	98.8	84.8	94.5	61.2	72.8	24.4	10.6	90.0	33.6	90.5	1.0	60.0	87.1	50.8
Qwen3-0.6B (generic)	64.2	74.0	67.5	68.6	83.0	98.8	95.3	94.0	61.4	71.2	22.8	8.3	89.3	32.2	88.8	1.1	61.1	86.8	51.2
jina-v3	55.8	32.8	43.3	73.4	78.5	98.7	38.0	93.4	62.6	73.4	19.9	0.7	90.8	39.2	77.7	0.6	73.0	89.1	18.6
snowflake-l-v2	58.4	22.8	59.1	74.0	78.5	98.7	77.3	93.8	66.5	73.1	20.3	5.7	86.9	19.1	83.6	1.0	44.5	90.5	55.0

Appendix

Table A8: Evaluation Results on MMTEB Retrieval Tasks (nDCG@10%)

Model	Avg	Alloprof	RuBQ	T2R	Voyage	WebLINX	Wiki-Multi
Qwen3-4B	65.08	85.13	72.28	67.27	65.61	11.30	88.89
jina-v4	62.20	78.24	70.95	64.86	58.86	14.00	86.28
Qwen3-0.6B (instr.)	61.41	80.38	65.67	67.15	57.66	11.60	85.99
Qwen3-0.6B (generic)	62.25	79.52	69.66	66.96	61.38	10.13	85.83
jina-v3	57.09	72.93	65.56	65.61	50.76	9.84	77.81
snowflake-l-v2	63.67	75.84	73.75	67.57	66.63	9.05	89.18

Appendix

Table A9: Evaluation Results on MMTEB Reranking Tasks (MAP@1000 [%]).

Model	Avg	Faro	FinP	GSTS	Indic	JSICK	SICK-R	S12	S13	S14	S15	S17	S22	STSB	STSBm	STSES	SRel
Qwen3-4B	80.9	85.8	34.0	90.0	60.8	88.8	88.1	86.6	94.4	90.9	93.8	91.8	73.0	86.1	93.7	72.8	63.2
jina-v4	74.4	72.3	15.1	88.2	35.2	80.3	89.2	83.5	88.6	84.8	89.7	85.0	71.8	86.6	88.6	75.3	56.5
Qwen3-0.6B (instr.)	76.2	74.3	26.3	84.9	39.0	86.6	84.8	83.0	91.8	87.1	91.4	85.5	71.8	84.6	91.1	76.9	59.4
Qwen3-0.6B (generic)	76.2	74.3	26.3	84.9	39.0	86.6	84.8	83.0	91.8	87.1	91.4	85.5	71.8	84.6	91.1	76.9	59.4
jina-v3	77.1	80.8	22.4	87.9	54.7	78.2	89.6	82.4	89.5	85.0	89.3	85.9	71.1	85.4	89.4	77.9	64.6
snowflake-l-v2	70.1	70.9	22.1	77.0	47.2	81.6	74.0	71.2	80.4	75.1	82.9	74.4	68.7	72.0	79.2	78.7	66.3

Appendix

Table A10: Evaluation Results on MMTEB Semantic Textual Similarity Tasks (Spearman correlation [%]).

Model	Avg	ArmPC	CTK	Opus	PawsX	Ppc	RTE3	Sprint	TERRa	TwURL	XNLI	IndoNLI
Qwen3-4B	85.1	96.3	86.2	95.8	68.9	95.3	90.8	96.1	66.6	87.2	87.2	65.3
jina-v4	79.3	94.4	79.6	94.1	61.9	91.8	88.3	91.4	57.4	86.4	71.9	54.6
Qwen3-0.6B (instr.)	80.8	93.1	79.2	92.9	62.2	90.5	89.0	94.1	60.7	86.7	81.7	59.0
Qwen3-0.6B (generic)	80.8	93.1	79.2	92.9	62.2	90.5	89.0	94.1	60.7	86.7	81.7	59.0
jina-v3	79.3	95.8	79.2	94.6	54.4	91.4	88.1	97.0	59.2	84.1	73.7	54.4
snowflake-l-v2	76.7	95.9	74.5	92.7	56.6	87.1	86.2	96.5	53.8	85.4	64.3	50.7

Appendix

Table A11: Evaluation Results on MMTEB Pair Classification Tasks (Max Average Precision [%]).

Model	Avg	BUCC	Bible	Bornh	DiaBl	Flores	IN22	Indic	NTX	Nolly	Norw	NusaT	NusaX	Tato
Qwen3-4B	79.4	98.9	25.9	71.3	87.1	74.1	82.0	93.7	87.8	63.6	93.1	91.1	86.9	76.3
jina-v4	62.4	98.5	11.3	34.0	84.8	53.5	68.0	79.3	71.7	32.1	93.1	63.4	64.6	57.2
Qwen3-0.6B (instr.)	72.2	98.4	21.2	54.2	80.9	62.8	75.8	90.4	79.5	61.3	92.7	85.4	78.4	58.1
Qwen3-0.6B (generic)	72.2	98.4	21.2	54.2	80.9	62.8	75.8	90.4	79.5	61.3	92.7	85.4	78.4	58.1
jina-v3	65.3	98.4	10.5	37.3	85.1	55.3	72.8	87.0	78.3	33.2	92.6	62.3	64.2	71.1
snowflake-l-v2	64.1	98.1	10.3	42.0	80.5	54.7	71.4	84.7	77.4	33.6	92.5	62.8	58.2	66.9

Appendix

Table A12: Evaluation Results on MMTEB Bitext Mining Tasks (F1 Score [%]).

Model	Afr	ACF	BgS	CSF	Cat	Cyr	CzP	DBP	Dal	Est	Fil	Fin	Grk	Guj	Ind	Idn	Zul	Ita	Kor	Kur	Mac	Mas
Qwen3-4B	50.9	92.5	74.9	58.7	52.6	95.0	68.3	99.3	50.7	59.6	47.8	93.9	51.9	92.0	95.1	64.3	23.2	66.7	61.1	81.9	75.5	84.3
jina-v4	40.9	72.0	61.1	31.6	49.5	34.2	52.3	81.0	49.9	39.8	30.7	78.2	26.0	82.7	20.2	60.0	22.5	58.7	57.0	65.7	53.5	72.4
Qwen3-0.6B (instr.)	46.0	90.4	73.4	44.1	49.1	90.2	63.8	98.8	50.0	41.1	40.8	90.6	39.7	85.8	93.1	64.8	25.6	61.7	55.9	80.1	61.0	80.5
Qwen3-0.6B (gen.)	40.0	70.1	65.4	32.3	47.9	74.3	50.9	96.8	50.0	34.6	28.6	74.6	39.3	81.3	36.5	59.0	23.6	70.1	58.1	71.1	47.2	77.0
jina-v3	42.8	92.2	76.0	39.1	45.1	43.3	61.9	76.1	50.1	48.8	38.5	79.5	11.6	86.2	18.0	58.1	20.8	62.8	56.2	57.6	61.2	71.5
snowflake-l-v2	44.1	64.7	59.3	31.0	49.1	57.7	49.7	89.6	50.2	40.7	31.8	72.0	35.7	84.4	33.0	60.9	23.9	68.1	58.4	61.7	61.3	74.3

Appendix

Table A13: Evaluation Results on MMTEB Classification Tasks (Accuracy [%]).

Model	Avg.	BrazilianToxic	CEDR	KorHate	MalteseNews	MultiEURLEX
Qwen3-4B	26.8	20.6	51.0	14.9	42.1	5.1
jina-v4	19.3	17.6	40.8	7.8	26.7	3.8
Qwen3-0.6B (instr.)	24.6	22.6	49.9	9.7	36.1	4.6
Qwen3-0.6B (generic)	21.1	23.7	38.7	8.8	29.7	4.4
jina-v3	18.4	19.7	47.4	10.6	12.7	1.5
snowflake-l-v2	18.9	22.5	38.5	10.5	18.7	4.6

Appendix

Table A14: Evaluation Results on MMTEB Multi-Label Classification Tasks (Accuracy [%]).

Model	Avg	Allo	AXP	AXS	BigPat	BioP	CLSP	HalS	MNC	MedP	PlscP	Rom	S200	SEx	SCP	Cities	WikiP
Qwen3-4B	57.2	59.5	64.8	65.2	43.2	50.9	73.2	30.5	56.2	45.3	75.1	44.3	41.3	77.6	62.1	93.6	31.6
jina-v4	44.0	44.5	57.9	57.6	28.7	38.4	37.4	25.1	40.5	35.2	69.6	41.1	27.4	54.0	27.3	92.0	26.8
Qwen3-0.6B (instr.)	52.3	54.0	63.7	63.8	32.5	47.3	62.0	29.0	53.2	42.2	74.2	40.3	34.1	71.2	53.4	86.8	29.7
Qwen3-0.6B (generic)	49.8	53.8	63.2	61.8	32.3	45.1	48.0	28.8	50.4	40.7	72.8	39.7	34.8	66.2	49.8	80.9	28.6
jina-v3	45.7	44.7	58.9	55.9	37.1	40.0	39.4	29.3	46.2	38.2	71.5	40.6	32.0	56.7	26.8	85.3	27.8
snowflake-l-v2	42.8	45.7	57.2	53.1	34.1	37.2	34.2	24.9	42.9	35.0	71.1	39.6	25.4	56.3	26.0	74.2	27.1

Appendix

Table A15: Evaluation Results on MMTEB Clustering Tasks (V-measure [%] - see footnote in Section 5.1 )

Model	Avg	Core17	News21	Robust04
Qwen3-4B	11.56	13.53	8.71	12.44
jina-v4	0.71	3.24	0.34	$-$ 1.46
Qwen3-0.6B (instr.)	5.09	8.93	3.61	2.75
Qwen3-0.6B (generic)	3.84	6.30	2.59	2.63
jina-v3	$-$ 1.34	$-$ 0.06	2.36	$-$ 6.31
snowflake-l-v2	$-$ 2.47	0.43	$-$ 0.24	$-$ 7.60

Appendix

Table A16: Evaluation Results on Instruction Reranking Tasks (p-MRR [%]).

Model	Avg	Arg	CQA	CF	DB	FEV	FiQA	HPQA	MSM	NFC	NQ	Quora	SCD	SCF	TREC	TOU
Qwen3-4B	61.6	75.6	50.3	47.4	48.2	91.6	62.7	74.7	42.7	41.1	63.1	88.1	31.4	78.3	92.9	35.4
jina-v4	54.0	67.0	43.7	35.1	43.9	87.9	48.5	68.5	38.1	34.4	61.7	78.6	21.5	76.1	80.4	24.1
Qwen3-0.6B	55.5	71.0	46.0	42.4	39.5	88.1	46.6	65.7	38.0	36.7	53.5	87.8	24.4	69.7	90.5	33.2
jina-v3	53.2	43.3	42.6	42.4	41.0	89.1	47.4	64.7	40.8	36.6	64.2	89.2	19.9	72.5	77.7	26.3
snowflake-l-v2	55.2	59.1	45.9	41.8	43.4	91.5	45.4	68.2	44.9	35.1	63.7	88.8	20.3	70.9	83.6	25.9
mult.-e5-l-instr.	52.7	58.5	44.3	29.9	38.4	78.0	48.4	69.3	40.4	36.3	57.8	89.2	19.2	71.6	82.5	27.4

Appendix

Table A17: Retrieval performance on BeIR (nDCG@10 [%]).

Model	Avg	NaQA	Needle*	Passkey*	QMSum	SummScreen	Wikim
Qwen3-4B	78.82	68.94	75.50	84.25	52.35	97.96	93.92
jina-v4	69.88	58.71	60.75	69.75	46.06	96.96	87.02
Qwen3-0.6B	72.20	63.25	50.75	84.75	47.70	96.72	90.00
jina-v3	55.67	34.30	64.00	38.00	39.34	92.33	66.02
snowflake-l-v2	63.74	43.63	50.25	77.25	40.08	96.38	74.84
mult.-e5-l-instr.	41.76	26.71	29.50	37.75	26.08	72.75	57.79

Appendix

Table A18: Retrieval performance on LongEmbed (nDCG@10 [%])

Model	Avg	ACD	AST	LS	LQA	FinB	HC3	FQA	HuE	MBPP	MIR	Apps	DS1K	WSQL	CDR	CURE	Fsh
Qwen3-4B	70.8	39.4	81.2	66.4	66.7	77.5	68.9	63.4	98.4	91.4	69.5	89.2	64.1	84.7	71.5	56.8	43.2
jina-v4	66.5	45.2	50.1	59.9	63.9	79.9	63.4	68.4	97.2	89.9	62.9	78.3	64.1	96.1	64.1	53.2	27.5
Qwen3-0.6B	64.2	36.1	79.0	63.6	53.4	74.8	54.5	56.3	92.3	87.0	61.2	75.3	59.7	86.8	62.5	47.0	37.9
jina-v3	54.6	34.8	32.8	59.2	59.4	72.2	61.5	39.2	80.6	83.2	62.6	29.0	50.1	68.0	64.2	46.3	30.5
snowflake-l-v2	54.0	34.0	22.8	66.2	65.5	75.6	54.4	56.6	71.5	80.2	66.5	9.7	42.7	69.7	60.7	54.7	32.3
mult.-e5-l-instr.	54.8	33.3	29.7	68.1	51.2	79.7	51.2	45.1	86.3	83.6	57.7	34.9	49.4	80.7	55.2	42.8	27.6

Appendix

Table A19: Retrieval performance on RTEB (Public) (nDCG@10 [%]).

Figure 6: Learning rate sensitivity across different optimization objectives. We report the average nDCG@10 on the MTEB (English, v2) benchmark using the S2ORC dataset. The plots compare $1\times 10^{-4}$ (blue) and $1\times 10^{-5}$ (orange) learning rates for embedding-based distillation ( $\mathcal{L}_{\text{distill}}$ ), InfoNCE ( $\mathcal{L}_{\text{NCE}}^{q\rightarrow d}$ ), and score-based distillation ( $\mathcal{L}_{\text{score}}$ ), all utilizing a trainable student projection.

Figure 7: Performance of Models on different languages on MMTEB compared to average performance

Text Embedding Models

Text embedding models map words, sentences, or documents to vectors for semantic search, retrieval, clustering, and downstream prediction.

See all 126 papers

Aug 24, 2026

Giga-Embeddings: Mixture-of-Experts Encoders for High-Throughput Text Embeddings

We introduce Giga-Embeddings, a family of text embedding models designed to combine strong retrieval quality with efficient serving. Its largest member is a sparse 10B-parameter Mixture-of-Experts encoder with approximately 1.8B active parameters per token. Across English, Russian, multilingual, and code MTEB benchmarks, this model achieves the strongest aggregate performance within the family on all four evaluated suites. In our vLLM benchmark with 1024-token inputs, it processes 114.5k tokens per second, providing 25 percent higher throughput than the dense 3B model and 1.56-2.65x the throughput of the evaluated external systems. The family also includes a dense 3B encoder and a distilled 480M encoder for tighter compute and memory budgets. We train the compact model using a dimension-agnostic objective that aligns teacher and student similarity distributions. The resulting 480M model scores 70.98 on Russian MTEB, surpassing FRIDA while using 42 percent fewer parameters. We release all three model checkpoints.

Egor Kolodin, Egor Krasnoperov, Evgeniy Kosarev +1

1MIPT · 2SaluteDevices · 3MSU

Text Embedding Models Mixture-of-Experts Models

Ask PDF

Mar 11, 2026

LLM2Vec-Gen: Generative Embeddings from Large Language Models

Fine-tuning LLM-based text embedders via contrastive learning maps inputs and outputs into a new representational space, discarding the LLM's output semantics. We propose LLM2Vec-Gen, a self-supervised alternative that instead produces embeddings directly in the LLM's output space by learning to represent the model's potential response. Specifically, trainable special tokens are appended to the input and optimized to compress the LLM's own response into a fixed-length embedding, guided by an unsupervised embedding teacher and a reconstruction objective. Crucially, the LLM backbone remains frozen and training requires only unlabeled queries. LLM2Vec-Gen achieves state-of-the-art self-supervised performance on the Massive Text Embedding Benchmark (MTEB), improving by 8.8% over the unsupervised embedding teacher. Since the embeddings preserve the LLM's response-space semantics, they inherit capabilities such as safety alignment (up to 22.6% reduction in harmful content retrieval) and reasoning (up to 35.6% improvement on reasoning-intensive retrieval). Finally, the learned embeddings are also interpretable: they can be decoded back into text to reveal their semantic content. Our code, data, and pretrained models are publicly available at https://github.com/McGill-NLP/llm2vec-gen.

Parishad BehnamGhader, Vaibhav Adlakha, Fabian David Schmidt +3

McGill University · Mila–Quebec AI Institute · ServiceNow Research +3

Text Embedding Models Self-Supervised Pre-Training

Ask PDF

Aug 13, 2026

The Embedder's Dilemma: LLMs Are Better, but at What Cost?

Should you replace your text-embedding pipeline with a large language model? We answer this with a controlled, cost-aware comparison of ten LLMs across six families and 26 embedding models (118M to 14B parameters) on 37 tasks spanning classification, semantic textual similarity (STS), clustering, pair classification, and retrieval. In aggregate the two paradigms are effectively tied: the best LLM (Gemini 3.1 Pro, 77.6) and the best embedding model (77.2) differ by 0.4 points. Their strengths differ by task: LLMs lead on reasoning-heavy retrieval, embedding models lead on classification, and the two match on clustering, STS, and pair classification. Reaching that parity is expensive. An LLM costs up to 1,431x more than an embedding model of comparable quality (USD 154 vs. USD 0.11 per benchmark pass), and the open LLMs tested process tokens 2.5 to 736x more slowly on the same GPU. Reasoning tokens account for 28 to 81% of LLM inference cost; lower reasoning budgets preserve or improve retrieval quality for most models in our ablation. The Pareto frontier contains the leading embedding models and one LLM, Gemini 3.1 Pro. These results support a division of labour: use embedding models for similarity, classification, and clustering, and reserve LLMs for reasoning-intensive retrieval. Our code, datasets, and results are publicly available at https://github.com/embeddings-benchmark/embedders-dilemma.

Adnan El Assadi, Niklas Muennighoff, Jinhyuk Lee

Harvard University · Stanford University · Independent Researcher

LLM Inference Efficiency LLM Evaluation Text Embedding Evaluation Information Retrieval Text Embedding Models

Ask PDF

Abstract

Text Embedding Models

Explore similar work

Giga-Embeddings: Mixture-of-Experts Encoders for High-Throughput Text Embeddings

LLM2Vec-Gen: Generative Embeddings from Large Language Models

The Embedder's Dilemma: LLMs Are Better, but at What Cost?