Citation and evidence

Best open-weight embedding models

38 min full readUpdated 102 references

This article's verification

Report a problem with this article

Edit article
More

Cite this article

Best open-weight embedding models is a comparison of text and multimodal embedding models, and of the rerankers used alongside them, whose trained weights can be downloaded and run on a user's own hardware, current to 19 November 2026. It covers about 60 model lines and groups them by the licence attached to the weights (permissive, conditional or non-commercial) and by documented use. It does not place them in one ranked order. Evidence of quality and adoption comes from the public leaderboard of the Massive Text Embedding Benchmark (MTEB) and from download counts on Hugging Face, each quoted with its date, the benchmark version and what the figure measures.[1][2]

On 19 November 2026 the MTEB(Multilingual, v2) leaderboard listed 492 models, 442 of them flagged as having open weights. The four highest-ranked entries all had downloadable weights, but under four different sets of terms: the MIT License (Microsoft's harrier-oss-v1-27b), a Tencent licence that says the model is not intended for use within the European Union (KaLM-Embedding-Gemma3-12B-2511), an NVIDIA licence limited to non-commercial research (llama-embed-nemotron-8b) and the Apache License 2.0 (Qwen3-Embedding-8B).[1][3][4][5][6] On the same day the most downloaded embedding repository on Hugging Face was all-MiniLM-L6-v2, a 22.7-million-parameter model under Apache 2.0 with 169.0 million downloads in the previous month. It ranked 152nd on that leaderboard.[2][7][1]

Every model here is called open-weight, not open source. The term describes the availability of the weights only. Whether a model also meets the Open Source AI Definition is a separate question that this article does not assess. Developers use their own wording, and it is not always consistent. The Qwen team's announcement calls the Qwen3 Embedding series "a new proprietary model of the Qwen model family" and, in the same post, says "We have open-sourced this series of text embedding and reranking models under the Apache 2.0 license"; the model card repeats the phrase "the latest proprietary model of the Qwen family" above an Apache 2.0 licence tag.[8][6] The card for nomic-embed-text-v2-moe calls the model "Fully Open-Source", and the card for Nomic's 3B multimodal model describes "Open-Weights: Model weights available for research use".[9][10] For generative models see Best open-source LLMs, and for licence texts see Open-weight license comparison.

How the models were selected

A model line is included only if its weights could be downloaded from Hugging Face and its licence could be read on 19 November 2026, in the model card, the repository's licence file or the Hub's licence tag. It also had to meet one of three tests.

TestCriterionSource of the figure
LeaderboardOne of the ten highest-ranked open-weight entries with a complete set of results on MTEB(Multilingual, v2)MTEB leaderboard API, read 2026-11-19 [1]
DownloadsA text embedding repository published by the model's own developer with at least 4 million downloads in the previous month, in the Hub's Sentence Similarity or Feature Extraction listingsHugging Face Hub API, read 2026-11-19 [2]
IllustrationA model that documents a licence tier or a capability the first two tests do not reach: multimodal input, late interaction, code retrieval, reranking or on-device useThe model's own card

Expanded article table

The third group is illustrative and not exhaustive, and inclusion under it is not a judgement of quality. Re-uploads and format conversions by third parties were not counted in the download test. The google/embeddinggemma-300m repository is gated and its card could not be read without signing in, so its details below come from Google's public model card and terms instead.[11][12]

Hugging Face counts a download each time a repository's query file (by default config.json) is requested over HTTP. The count does not deduplicate users and includes automated pipelines, so it measures how often a repository is fetched, not how many people or products rely on it.[13]

Reading MTEB scores

MTEB was introduced in 2022 by Niklas Muennighoff and co-authors. The current version of the paper describes 8 embedding tasks over 58 datasets and 112 languages; the first arXiv version, of 13 October 2022, counted 56 datasets. Its authors reported that "no particular text embedding method dominates across all tasks".[14] MMTEB, published in 2025, extended it to "over 500 quality-controlled evaluation tasks across 250+ languages".[15] The leaderboard now hosts many named benchmarks, each with its own task list and version number.[1]

Benchmark (as named on the leaderboard)TasksWhat it measures, per the leaderboardVersion note
MTEB(Multilingual, v2)131Text embedding quality across 250+ languages: bitext mining, classification, clustering, retrieval, reranking, semantic similarityReplaced v1 (132 tasks) after one dataset was removed from the Hugging Face Hub
MTEB(eng, v2)41English classification, clustering, pair classification, reranking, retrieval, semantic similarity, summarizationReplaced v1 (56 tasks)
MTEB(Code, v1)12Code retrieval: code search, text-to-SQL, code feedback
RTEB(beta)30Retrieval in legal, finance, code and healthcare domains, on open and closed datasetsBeta
ViDoRe(v3)10Retrieval over images of enterprise documents, on open and closed datasets

Expanded article table

All rows are from the leaderboard's benchmark list as read on 19 November 2026.[1]

Scores from different versions are not comparable

The leaderboard recommends MTEB(eng, v2) over v1 because v2 "resolves a known scoring bug, uses updated task versions, and removes common fine-tuning datasets such as MSMARCO for more comparable scores".[1] A figure from one version therefore cannot be set beside a figure from the other. NVIDIA's card for NV-Embed-v2 reports a score of 72.31 "across 56 text embedding tasks" as of 30 August 2024, which is the v1 task count; on MTEB(eng, v2) the same model had a Mean(Task) of 69.81 and ranked 26th on 19 November 2026.[16][1]

Model cards also freeze a ranking at the date they were written. The Qwen3-Embedding-8B card says the model "ranks No.1 in the MTEB multilingual leaderboard (as of June 5, 2025, score 70.58)". Seventeen months later the same 70.58 placed fourth.[6][1]

Rank and mean are different measures

The leaderboard's rank column is Rank (Borda). The maintainers describe it as a Borda count in which "each task is treated as a preference voter", and note that it "tends to prefer models that perform well broadly across tasks". They describe Mean(Task) as "a naïve average computed across all the tasks within the benchmark" that "can overvalue tasks with higher variance in its scores".[17] The two can disagree: on MTEB(Multilingual, v2), llama-embed-nemotron-8b ranked third with a Mean(Task) of 69.46, ahead of Qwen3-Embedding-8B at 70.58.[1]

Zero-shot share and who reports the scores

Each row carries a zero-shot percentage. The maintainers define a model as zero-shot "if it is not trained on any splits of the datasets used to derive the tasks". A share of 100% means none of a benchmark's datasets were used in training. At a share of 50%, they write, "the benchmark results should be interpreted with a pinch of salt". Derivative models inherit the training datasets of their parents.[17] The share is computed from the training datasets declared in each model's metadata, and the documentation asks users who find undeclared overlap to report it.[18] For some rows the leaderboard's API returns -1 instead of a percentage; on 19 November 2026 that was the case for 93 of the 492 MTEB(Multilingual, v2) rows, among them the Jina Embeddings v5 rows, nomic-embed-text-v2-moe and gte-small.[1]

Results reach the leaderboard through pull requests to a public results repository, which anyone can open. The documentation says that reviewers "might ask you about suspicious results or ask you to check for potential data leakage".[19] The 2025 announcement of RTEB, written by four authors from the MTEB organisation and two from MongoDB, put the underlying problem this way: "When training data sources overlap with evaluation datasets, a model's score can become inflated". RTEB responds by mixing open datasets with private ones that only the maintainers evaluate, and the announcement adds that "models with slightly lower benchmark performance and a higher zero-shot score are often recommended instead".[20] See also data contamination.

A score printed on a model card is the developer's own run and can differ from the leaderboard's. Google's card for EmbeddingGemma 2 gives 61.36 on MTEB(Multilingual, v2); the leaderboard showed 60.52 for the same model on 19 November 2026.[21][1]

What the maintainers say a leaderboard misses

In a June 2026 post about the rebuilt leaderboard, three MTEB maintainers wrote that a benchmark "rarely measures what you care about" and that a predefined leaderboard "might only contain 50% of the tasks that you care about". They also wrote that ranking by performance encourages development "toward better performance while ignoring other factors such as size, memory usage, and runtime".[22]

Leaderboard snapshot: MTEB(Multilingual, v2)

The table lists the ten highest-ranked open-weight entries that had results for all 131 tasks on 19 November 2026, in the leaderboard's own Borda order. Ranks count every row, including closed models and rows with a low zero-shot share. Of the 492 rows, 96 had a complete set of results and so a mean. Parameter counts, dimensions and input limits are those shown by the leaderboard, except for harrier-oss-v1, where the model card's input limit is used.[1][3]

Rank (Borda)ModelDeveloperWeights licenceParametersDimensionsMax input tokensZero-shotMean(Task)Mean(TaskType)
1harrier-oss-v1-27bMicrosoftMIT [3]27.0B5,37632,76878%74.2764.20
2KaLM-Embedding-Gemma3-12B-2511TencentTencent KaLM licence (not intended for use in the EU) plus Gemma terms [4]11.8B3,84032,76873%72.3262.51
3llama-embed-nemotron-8bNVIDIANVIDIA License (non-commercial) plus Llama 3.1 licence [5]7.5B4,09632,76899%69.4661.09
4Qwen3-Embedding-8BAlibaba (Qwen)Apache 2.0 [6]7.6B4,09632,76899%70.5861.69
6Qwen3-Embedding-4BAlibaba (Qwen)Apache 2.0 [6]4.0B2,56032,76899%69.4560.86
7Octen-Embedding-8BOctenApache 2.0 [23]7.6B4,09632,76899%67.8460.28
8F2LLM-v2-14BCodeFuseApache 2.0 [24]14.0B5,12040,96088%68.7459.45
9F2LLM-v2-8BCodeFuseApache 2.0 [24]7.6B4,09640,96088%68.0958.99
10harrier-oss-v1-0.6bMicrosoftMIT [3]0.6B1,02432,76878%69.0159.00
12F2LLM-v2-4BCodeFuseApache 2.0 [24]4.0B2,56040,96088%67.0658.25

Expanded article table

Ranks 5 and 11 were closed models.[1] Octen-Embedding-8B is a fine-tune of Qwen3-Embedding-8B.[23] Microsoft's card says the 270m and 0.6b harrier models were also trained by knowledge distillation from larger embedding models.[3]

Smaller models placed lower on the same board: Qwen3-Embedding-0.6B was 20th (Mean(Task) 64.34), multilingual-e5-large-instruct 27th (63.22), EmbeddingGemma 31st (61.14), EmbeddingGemma 2 40th (60.52), Jina Embeddings v3 51st (58.37), nomic-embed-text-v2-moe 57th (57.64) and snowflake-arctic-embed-l-v2.0 64th (57.03).[1]

On RTEB(beta), which scores retrieval only, Qwen3-Embedding-8B ranked first on 19 November 2026 with a mean of 69.37. The leaderboard notes that it has temporarily removed the benchmark's column of private-dataset scores, and its API returned a mean for only 23 of the 300 RTEB rows that day.[1] The Octen-Embedding-8B card claims first place on RTEB "as of January 12, 2026"; the leaderboard showed it fifth, without a mean, on 19 November 2026.[23][1]

Most-downloaded embedding models

The table lists the sixteen repositories that met the download test, in order of downloads. Two leaderboard columns give each model's rank and Mean(Task) on 19 November 2026; "no mean" marks a row that lacked results for some tasks. The BGE v1.5 English models, all-MiniLM-L6-v2, all-mpnet-base-v2, nomic-embed-text-v1.5, gte-small and the Granite English model are English models, and their multilingual figures appear because the leaderboard reports them, not because the models claim multilingual coverage.[2][1]

ModelDeveloperWeights licenceDownloads, previous month (millions)ParametersDimensionsMax inputMTEB(eng, v2)MTEB(Multilingual, v2)
all-MiniLM-L6-v2Sentence TransformersApache 2.0 [7]169.0322.7M384256 word pieces151st, 59.03152nd, 41.39
bge-small-en-v1.5BAAIMIT [25]43.9633.4M384512 tokens89th, 64.30122nd, 43.76
bge-m3BAAIMIT [26]31.74568M1,0248,192 tokens225th, no mean52nd, no mean
paraphrase-multilingual-MiniLM-L12-v2Sentence TransformersApache 2.0 [27]25.22117.7M384128 (card); 512 (leaderboard)195th, 56.42104th, 48.73
all-mpnet-base-v2Sentence TransformersApache 2.0 [28]19.90109.5M768384 word pieces128th, no mean139th, no mean
nomic-embed-text-v1.5Nomic AIApache 2.0 [29]11.52136.7M7688,192 tokens123rd, 62.20117th, 44.10
multilingual-e5-smallintfloatMIT [30]9.45117.7M384512 tokens161st, 59.6973rd, no mean
paraphrase-multilingual-mpnet-base-v2Sentence TransformersApache 2.0 [31]9.38278.0M768128 (card); 512 (leaderboard)158th, 58.2783rd, 51.93
bge-base-en-v1.5BAAIMIT [25]8.60109.5M768512 tokens77th, 65.14111th, 44.52
Qwen3-Embedding-0.6BAlibaba (Qwen)Apache 2.0 [6]7.89595.8M1,02432,768 tokens28th, 70.4720th, 64.34
multilingual-e5-largeintfloatMIT [32]7.33559.9M1,024512 tokens128th, no mean54th, no mean
bge-large-en-v1.5BAAIMIT [25]7.05335.1M1,024512 tokens69th, 65.8998th, 45.08
granite-embedding-small-english-r2IBMApache 2.0 [33]6.6647.7M3848,192 tokens132nd, 61.07243rd, no mean
multilingual-e5-baseintfloatMIT [34]4.70278.0M768512 tokens146th, 61.0069th, 57.02
snowflake-arctic-embed-l-v2.0SnowflakeApache 2.0 [35]4.64567.8M1,0248,192 tokens104th, 63.6064th, 57.03
gte-smallAlibaba DAMO AcademyMIT [36]4.0633.4M384512 tokens96th, 63.22123rd, no mean

Expanded article table

Parameter counts are the safetensors totals reported by the Hub, except for bge-m3, whose repository reports none and for which the leaderboard's figure is used. The 512-token limits for the multilingual E5 models and gte-small are the ones stated on their cards; the leaderboard lists 514 for multilingual-e5-base and multilingual-e5-large. For the two paraphrase-multilingual models the sources disagree: the leaderboard lists a 512-token limit for both, while the architecture printed on each card shows a maximum sequence length of 128, so the table gives both figures.[2][1][27][31][30][32][34][36]

Every model in this table carries an MIT or Apache 2.0 licence tag. Lower in the same listings, Jina Embeddings v3 had 1.36 million downloads under a non-commercial licence and the gated EmbeddingGemma had 1.02 million under the Gemma terms.[2][37][11]

Licence tiers

The licence on the weights sets the terms under which a downloaded model may be used. The three tiers below follow the wording of each card or licence file: permissive where the weights carry no use restriction beyond notice and attribution, conditional where commercial use is allowed subject to further terms, and non-commercial where the card or licence limits use to research or non-commercial purposes. Background on the licence families is in Open-source license and Creative Commons licenses.

Permissive: Apache 2.0, MIT and OpenMDW

Dates are the developer's stated release date where the card or announcement gives one; "repo" marks the date the Hugging Face repository was created instead. The Hub records no creation date earlier than 2 March 2022, so for older repositories the date of the first commit is given.[100]

Model lineDeveloperWeights licenceReleasedParametersDimensionsMax input tokensLanguages, per the card
Qwen3 Embedding 0.6B, 4B, 8BAlibaba (Qwen)Apache 2.0 [6][8]2025-06-050.6B, 4B, 8B1,024, 2,560, 4,09632kOver 100, plus programming languages
harrier-oss-v1 270m, 0.6b, 27bMicrosoftMIT [3]repo 2026-03-30270M, 0.6B, 27B640, 1,024, 5,37632,768Multilingual
EmbeddingGemma 2Google DeepMindApache 2.0 [21][38]2026-10-06740M (270M text-only)7688,192Over 100
BGE-M3BAAIMIT [26]2024-01-30 [25]568M1,0248,192Over 100
BGE v1.5 English: small, base, largeBAAIMIT [25]repo 2023-0933.4M, 109.5M, 335.1M384, 768, 1,024512English
Multilingual E5: small, base, large, large-instructintfloatMIT [32][39]repo 2023-05 to 2024-02117.7M, 278.0M, 559.9M, 559.9M384, 768, 1,024, 1,024512100, per the card of the large model
gte-multilingual-baseAlibabaApache 2.0 [40]repo 2024-07-20305M7688,192Over 70
gte-modernbert-baseAlibaba (Tongyi Lab)Apache 2.0 [41]repo 2025-01-20149M7688,192English
nomic-embed-text-v1.5Nomic AIApache 2.0 [29]repo 2024-02-10136.7M7688,192English
nomic-embed-text-v2-moeNomic AIApache 2.0 [9]repo 2025-02-07475M (305M active)768512About 100
snowflake-arctic-embed-l-v2.0SnowflakeApache 2.0 [35]2024-12-04568M1,0248,192Multilingual
mxbai-embed-large-v1MixedbreadApache 2.0 [42]repo 2024-03-07335.1M1,024512English
Stella v5: stella_en_400M_v5, stella_en_1.5B_v5NovaSearchMIT [43]repo 2024-07-12435M, 1.54B512 to 8,192 (default 1,024)8,192 on the leaderboard (400M model); the card recommends 512English
Granite Embedding R2: small English, English, 311M multilingualIBMApache 2.0 [33][44][45]2025-08-15 (English), 2026-04-29 (multilingual)47M, 149M, 311M384, 768, 7688,192, 8,192, 32,768English; over 200 for the multilingual model
all-MiniLM-L6-v2, all-mpnet-base-v2Sentence TransformersApache 2.0 [7][28]first commit 2021-08-3022.7M, 109.5M384, 768256 and 384 word piecesEnglish
paraphrase-multilingual-MiniLM-L12-v2, paraphrase-multilingual-mpnet-base-v2Sentence TransformersApache 2.0 [27][31]first commit 2021-06-02117.7M, 278.0M384, 768128 on the cards; 512 on the leaderboardMultilingual
gte-smallAlibaba DAMO AcademyMIT [36]repo 2023-07-2733.4M384512English
voyage-4-nanoVoyage AI (MongoDB)Apache 2.0 [46]repo 2026-01-06346M2,048, 1,024, 512 or 25632,000Multilingual
pplx-embed-v1 0.6b, 4bPerplexityMIT [47]repo 2026-01-14 (0.6b), 2026-01-20 (4b)0.6B, 4B1,024, 2,56032KMultilingual
Octen-Embedding-8BOctenApache 2.0 [23]repo 2025-12-237.6B4,09632,768Over 100
F2LLM-v2 (eight sizes)CodeFuseApache 2.0 [24]repo 2026-0380M to 14BUp to 5,12040,960Over 200
Nemotron-3-Embed-8BNVIDIAOpenMDW 1.1 [48]2026-07-167.95B4,09632,768Multilingual

Expanded article table

Dimensions and input limits for F2LLM-v2, Nemotron-3-Embed-8B, mxbai-embed-large-v1 and Stella, and the BGE-M3 parameter count, are the leaderboard's.[1] The Stella card says of sequence length that "512 is recommended" and that the model was trained on inputs of that length.[43]

Some cards go beyond the licence name. BAAI's cards state that "The released models can be used for commercial purposes free of charge", and Google's announcement describes EmbeddingGemma 2 as "released under a commercially permissive Apache 2.0 license".[25][38] NVIDIA's card for Nemotron-3-Embed-8B says the model is ready for commercial use and was built on an Apache 2.0 Ministral base; its OpenMDW 1.1 licence grants permission "to deal in the Model Materials without restriction", subject to compliance with the agreement.[48][49] A permissive licence on the weights says nothing about the terms of the data the model was trained on, which is covered under training data below.

Conditional: commercial use with conditions

ModelDeveloperTermsConditions stated
EmbeddingGemma (embeddinggemma-300m)Google DeepMindGemma Terms of Use [12][50]Gated on Hugging Face. Redistribution must pass the use restrictions on to recipients, and uses in Google's prohibited use policy are barred
bge-multilingual-gemma2BAAIGemma licence tag on the Hub [51]Gemma terms apply
llama-nemotron-embed-1b-v2NVIDIANVIDIA Open Model License plus the Llama 3.2 community licence [52]Card states the model is ready for commercial use; two licences apply together
llama-nemotron-rerank-1b-v2NVIDIAOpenMDW 1.1 plus the Llama 3.2 community licence [53]Two licences apply together
KaLM-Embedding-Gemma3-12B-2511TencentTencent KaLM licence plus Gemma terms for the base model [4][54]"KaLM-Embedding IS NOT INTENDED FOR USE WITHIN THE EUROPEAN UNION."
ColPali v1.3ViDoRe (ColPali authors)Gemma licence on the PaliGemma backbone, MIT on the adapters [55]The Hub tag reads MIT, which covers the adapters only

Expanded article table

The 300-million-parameter EmbeddingGemma of September 2025 remains under the Gemma terms, while its successor EmbeddingGemma 2 of October 2026 was released under Apache 2.0. They are different models with different weights, not one model relicensed.[12][38]

Non-commercial and research-only

ModelDeveloperTermsWhat the card or licence says
Jina Embeddings v3Jina AICC BY-NC 4.0 [37]Listed on AWS and Azure; for other commercial use the card says to contact the company
Jina Embeddings v5 text-smallJina AICC BY-NC 4.0 [56]"For commercial use, please contact us"
Jina Embeddings v4Jina AIQwen Research License [57]Licence grants use for non-commercial purposes only
jina-reranker-v3, jina-colbert-v2, jina-clip-v2, jina-code-embeddingsJina AICC BY-NC 4.0 [58][59][60][61]jina-clip-v2 is "available for commercial use" through Jina's API and cloud marketplaces
NV-Embed-v2NVIDIACC BY-NC 4.0 [16]"This model should not be used for any commercial purpose."
llama-embed-nemotron-8bNVIDIANVIDIA License (non-commercial) plus the Llama 3.1 community licence [5]"This model is for non-commercial/research use only"
SFR-Embedding-MistralSalesforceCC BY-NC 4.0 [62]Research purposes only
Linq-Embed-MistralLinq AI ResearchCC BY-NC 4.0 [63]Research purposes only
colnomic-embed-multimodal-3bNomic AIQwen Research License file, no Hub licence tag [10]"Model weights available for research use"
ColQwen2.5 v0.2ViDoRe (ColPali authors)Qwen Research License on the backbone, MIT on the adapters [64]The Hub tag reads MIT, which covers the adapters only
Reason-ModernColBERTLightOnCC BY-NC 4.0 [65]Non-commercial because of its training data

Expanded article table

For commercial use NVIDIA's NV-Embed-v2 card recommends the models in its NeMo Retriever microservices (NIMs) instead.[16]

Licence changes

DateModelChangeEvidence
2025-08-19jina-embeddings-v4Licence file added naming the Qwen Research License. The card says the model "was initially released under cc-by-nc-4.0 due to an error" and that the Qwen licence applies because the model derives from Qwen2.5-VL 3BCard and repository history [57]
2026-07-24zerank-1, zerank-2 (ZeroEntropy rerankers)CC BY-NC 4.0 to Apache 2.0Repository history: commits titled "Relicense model under Apache 2.0" [66]
2026-08-26llama-nemotron-rerank-1b-v2NVIDIA Open Model License to OpenMDW 1.1; the Llama 3.2 community licence still appliesRepository history: commit titled "Update license to OpenMDW-1.1" [53]

Expanded article table

Models by use case

The groupings below repeat what developers document. They are not rankings.

Small, CPU and on-device models

The most downloaded models are small: all-MiniLM-L6-v2 has 22.7 million parameters, bge-small-en-v1.5 and gte-small have 33.4 million each, and IBM's granite-embedding-small-english-r2 has 47 million.[7][25][36][33] Google states that EmbeddingGemma, at 308 million parameters, is "small enough to run on less than 200MB of RAM with quantization" and reports an inference time under 15 ms for 256 input tokens on EdgeTPU hardware.[67] For EmbeddingGemma 2 Google reports that, with quantization, the model needs about 191 MB of active RAM for text-only weights and about 567 MB for the full multimodal model on a Pixel 11 Pro.[38] Other small permissive entries are harrier-oss-v1-270m (MIT) and voyage-4-nano (Apache 2.0, 346 million parameters).[3][46] These memory and latency figures are the developers' own and were measured on the hardware each names.

Multilingual retrieval

ModelLanguage coverage stated by the developerLicence
BGE-M3More than 100 working languages [26]MIT
Multilingual E5The 100 languages of XLM-RoBERTa, with the caveat that "low-resource languages may see performance degradation" [32]MIT
Qwen3 EmbeddingOver 100 languages [6]Apache 2.0
gte-multilingual-baseOver 70 languages [40]Apache 2.0
EmbeddingGemma, EmbeddingGemma 2Over 100 languages [12][21]Gemma terms; Apache 2.0
granite-embedding-311m-multilingual-r2Over 200 languages, with enhanced support for 52 [45]Apache 2.0
nomic-embed-text-v2-moeAbout 100 languages [9]Apache 2.0
F2LLM-v2More than 200 languages [24]Apache 2.0
Jina Embeddings v3The card says the base model covers 100 languages and tuning focused on 30 [37]; Jina's announcement says 89 languages in total, "including 30 languages with the best performance" [101]CC BY-NC 4.0

Expanded article table

The MMTEB paper reported in February 2025 that the best-performing publicly available model on its benchmark was multilingual-e5-large-instruct, "with only 560 million parameters".[15] On 19 November 2026 that model ranked 27th on MTEB(Multilingual, v2).[1]

Long inputs

Input limits vary by a factor of more than a hundred. all-MiniLM-L6-v2 truncates input beyond 256 word pieces, and the multilingual E5 card says that "Long texts will be truncated to at most 512 tokens".[7][32] Google's card gives EmbeddingGemma a maximum input context of 2K tokens, which the leaderboard lists as 2,048.[12][1] BGE-M3, nomic-embed-text-v1.5, the GTE models listed above, snowflake-arctic-embed-l-v2.0, the English Granite R2 models and EmbeddingGemma 2 accept 8,192.[26][29][40][35][44][21] Qwen3 Embedding, harrier-oss-v1, pplx-embed-v1, Octen-Embedding-8B, Jina Embeddings v4 and the multilingual Granite R2 model document limits of about 32,000 tokens.[6][3][47][23][57][45] Splitting longer documents before embedding is covered in chunking.

Code retrieval

MTEB(Code, v1) has 12 tasks. On 19 November 2026 its highest-ranked row, papr-embed-code-1-4b, had a zero-shot share of 16%, and second-placed F2LLM-v2-14B had 58%. Rows with a 100% share included Qwen3-Embedding-4B at 5th (mean 80.07), Qwen3-Embedding-8B at 7th (80.69), pplx-embed-v1-4b at 12th (78.73), EmbeddingGemma 2 at 19th (77.88) and Qwen3-Embedding-0.6B at 32nd (75.42).[1] Models built for code include nomic-embed-code, a 7-billion-parameter Apache 2.0 model covering Python, Java, Ruby, PHP, JavaScript and Go, and jina-code-embeddings, which is CC BY-NC 4.0.[68][61]

Multimodal and document-image retrieval

Some models embed images, audio or video into the same space as text. EmbeddingGemma 2 takes text, images, video and audio (Apache 2.0).[21] Qwen3-VL-Embedding, in 2B and 8B sizes, takes text, images, screenshots and video (Apache 2.0).[69] Jina Embeddings v4 takes text, images and visual documents under the research-only Qwen licence.[57] Tencent's WeMM-Embedding 9B takes text, images, videos and visual documents and returns a 4,096-dimensional embedding; its licence file is Apache 2.0 and adds that third-party components remain under their original licences.[70] The SigLIP 2 checkpoint cited here carries an Apache 2.0 licence tag.[71] OpenAI's CLIP ViT-L/14 repository on Hugging Face carries no licence tag, although OpenAI's GitHub repository for CLIP is under the MIT License; the model card states that "Any deployed use case of the model - whether commercial or not - is currently out of scope".[72][102]

A second family retrieves pages as images. ColPali, introduced in 2024, uses a vision-language model to produce multi-vector embeddings directly from images of document pages, and its paper introduced the ViDoRe benchmark.[73] On ViDoRe(v3) on 19 November 2026 the highest-ranked row was tencent/EVIE-8B (mean 66.24), a model this article does not cover. Rows for models in this article were Qwen3-VL-Embedding-8B (23rd, mean 57.78), colnomic-embed-multimodal-7b (25th, 57.33), Jina Embeddings v4 (27th, 56.34, with a second row for the same model at 35th), colnomic-embed-multimodal-3b (29th, 55.78), Qwen3-VL-Embedding-2B (31st, 52.64), ColQwen2.5 v0.2 (32nd, 51.90), ColQwen2 v1.0 (38th, 44.65) and ColPali v1.3 (39th, 43.05).[1] Licences split within one family: colnomic-embed-multimodal-7b is tagged Apache 2.0, while the 3b version ships a research licence; ColQwen2 v1.0 has an Apache 2.0 backbone, while ColQwen2.5 has a research-licence backbone.[74][10][75][64] The ColPali v1.3 card says it was trained on English data only.[55]

Late-interaction models

ColBERT (2020) encodes the query and the document separately into one vector per token and compares them at search time, a design its authors call late interaction. It lets document representations be computed in advance.[76][77] Open-weight late-interaction models include ColBERTv2 (MIT), GTE-ModernColBERT-v1 from LightOn (Apache 2.0, 128-dimensional token vectors), answerai-colbert-small-v1 (Apache 2.0, 33 million parameters), mxbai-edge-colbert-v0-32m (Apache 2.0, 32 million parameters) and jina-colbert-v2 (CC BY-NC 4.0).[77][78][79][80][59] BGE-M3 produces dense, sparse and multi-vector output from one model, and its card recommends hybrid retrieval followed by re-ranking.[26]

Rerankers

Sentence Transformers describes Cross-Encoder models, which it also calls rerankers, as models that calculate similarity scores.[81] The BGE-M3 card recommends re-ranking as a step after retrieval.[26]

RerankerDeveloperWeights licenceParametersDownloads, previous monthNotes from the card
cross-encoder/ms-marco-MiniLM-L6-v2Sentence TransformersApache 2.0 [82]22.7M56.93MTrained on MS MARCO passage ranking
bge-reranker-v2-m3BAAIApache 2.0 [83]567.8M14.83MMultilingual
Qwen3-Reranker 0.6B, 4B, 8BAlibaba (Qwen)Apache 2.0 [84]0.6B, 4B, 8B1.16M, 2.87M, 0.06M32k context, over 100 languages
gte-reranker-modernbert-baseAlibabaApache 2.0 [85]149.6M1.38M8,192 tokens
mxbai-rerank-base-v2, mxbai-rerank-large-v2MixedbreadApache 2.0 [86]494M, 1.54B0.02M, 0.01MOver 100 languages
zerank-2ZeroEntropyApache 2.0 since 2026-07-24 [66]4.0B0.05MPreviously non-commercial
llama-nemotron-rerank-1b-v2NVIDIAOpenMDW 1.1 plus Llama 3.2 licence [53]1.24B0.06MSee licence changes
jina-reranker-v2-base-multilingualJina AICC BY-NC 4.0 [87]278.4M1.11M
jina-reranker-v3Jina AICC BY-NC 4.0 [58]596.8M0.74MListwise, up to 64 documents per call

Expanded article table

Download counts were read from the Hub on 19 November 2026.[2]

Dimension, size and accuracy

Matryoshka Representation Learning (MRL), published in 2022, trains a model so that a single embedding "encodes information at different granularities" and can be shortened to fit "the computational constraints of downstream tasks".[88] Several models in this article were trained with it, and some cards publish what truncation costs. The figures below are each developer's own measurements, and the three rows use different benchmarks.

ModelBenchmark named by the card76851225612864
EmbeddingGemma 2 [21]MTEB(Multilingual, v2), Mean(Task)61.3661.1760.4157.89
EmbeddingGemma [12]MTEB(Multilingual, v2), Mean(Task)61.1560.7159.6858.23
nomic-embed-text-v1.5 [29]"MTEB", no version named62.2861.9661.0459.3456.10

Expanded article table

Google summarises its own table as showing "minimal quality impact down to 256d".[21] The Stella v5 card, whose models offer dimensions from 512 to 8,192, states: "The higher the dimension, the better the performance."[43] Snowflake's card says that MRL and quantization-aware embedding training allow snowflake-arctic-embed-l-v2.0 to use "embeddings as small as 128 bytes/vector".[35]

Parameter count shows the same trade-off within a family on MTEB(Multilingual, v2) as read on 19 November 2026.[1]

FamilySizeDimensionsRank (Borda)Mean(Task)
harrier-oss-v1270M6401966.55
harrier-oss-v10.6B1,0241069.01
harrier-oss-v127B5,376174.27
Qwen3 Embedding0.6B1,0242064.34
Qwen3 Embedding4B2,560669.45
Qwen3 Embedding8B4,096470.58

Expanded article table

Cards also tie speed claims to architecture. Alibaba's card for gte-multilingual-base, an encoder-only model with 305 million parameters, claims a tenfold increase in inference speed over a decoder-based model of its own, gte-Qwen2-1.5B-instruct.[40]

Training data and its terms

The terms of a training dataset are separate from the licence on a model's weights, and cards handle the difference in different ways. One example is MS MARCO, whose site, hosted by Microsoft, says the datasets "are intended for non-commercial research purposes only".[89]

ModelWhat the card says about training data
all-MiniLM-L6-v2 (Apache 2.0)Lists 9,144,553 MS MARCO triplets among more than a billion training pairs [7]
multilingual-e5-large (MIT)Lists 500k MS MARCO pairs among its supervised fine-tuning data [32]
ColBERTv2 (MIT), cross-encoder/ms-marco-MiniLM-L6-v2 (Apache 2.0)Trained on MS MARCO passage ranking [77][82]
Granite Embedding R2 (Apache 2.0)"we do not use the popular MS-MARCO retrieval dataset in our training corpus due to its non-commercial license" [44]
llama-nemotron-embed-1b-v2 (NVIDIA Open Model License)Says MS MARCO "restricts commercial licensing" and that NVIDIA built its own blend from public datasets licensed for commercial applications [52]
Reason-ModernColBERT (CC BY-NC 4.0)"since the ReasonIR data has been released under a cc-by-nc-4.0 license, we cannot release this model under an Apache 2.0 license" [65]
mxbai-embed-large-v1 (Apache 2.0)"Our model was trained with no overlap of the MTEB data" [42]
Stella v5 (MIT)Training method and datasets "will be released in the future" [43]

Expanded article table

Software for running the models

All of the tools below ship under an OSI-approved licence, confirmed from the licence file in each repository. Repository figures are as of November 2026.

ToolMaintainerLicenceLatest releaseStarsWhat its documentation says
Sentence TransformersHugging Face (created by UKP Lab)Apache 2.0v6.3.0, 2026-11-1119,289Python library for using and training embedding, reranker, sparse and multi-vector models [81][90]
Text Embeddings Inference (TEI)Hugging FaceApache 2.0v1.9.4, 2026-09-155,128Server for embedding and sequence-classification (reranking) models, with CPU and CUDA Docker images, Metal support on Macs and an OpenAI-compatible route [91]
OllamaOllamaMITv0.40.16, 2026-11-09187,552/api/embed endpoint that returns L2-normalized vectors [92][93]
llama.cppggml-orgMITv1.0.0, 2026-11-13135,610Server flags --embedding and --reranking, with /v1/embeddings and /v1/rerank endpoints [94][95]
vLLMvLLM projectApache 2.0v0.34.0, 2026-11-1697,184Pooling models for embedding, classification and scoring [96][97]
Infinitymichaelfeil (GitHub account)MITv3.0.0, 2026-10-262,987REST server for embeddings, reranking and classification; v3 is written in Rust on LibTorch [98]
FastEmbedQdrantApache 2.0v0.9.1, 2026-10-243,303Python library on ONNX Runtime with no PyTorch dependency; dense, sparse, late-interaction and image embeddings, plus rerankers [99]

Expanded article table

Text Embeddings Inference was not always under Apache 2.0. Its first release in October 2023 used the Hugging Face Optimized Inference License 1.0, and the repository switched to Apache 2.0 on 8 April 2024. As of November 2026 the repository is not archived, and its latest release is dated 15 September 2026.[91]

The vLLM documentation carries a caution about its own embedding support: "We currently support pooling models primarily for convenience. This is not guaranteed to provide any performance improvements over using Hugging Face Transformers or Sentence Transformers directly."[96] The llama.cpp server documentation says that --embedding should be used "only with dedicated embedding models" and names bge-reranker-v2-m3 as an example model for its reranking endpoint.[94] Ollama's documentation advises readers to "Use the same embedding model for both indexing and querying."[92] TEI and the llama.cpp server both document an OpenAI-compatible embeddings route.[91][94] For the wider setup see Local AI and Self-hosting LLMs.

Limits and cautions

  • Licence tags can understate the terms. The Hub tag for ColQwen2.5 v0.2 and ColPali v1.3 reads MIT, but each card says the tag covers the adapters and that the backbone has its own licence. Jina Embeddings v4 was first published under the wrong licence by its developer's account, and Nomic's colnomic-embed-multimodal-3b repository has no licence tag at all.[64][55][57][10]
  • Models built on other models can carry two licences. NVIDIA's Llama-based embedders add a Llama community licence to NVIDIA's own, and KaLM-Embedding-Gemma3-12B-2511 adds the Gemma terms of its base model.[5][52][54]
  • Leaderboard position is not a measurement of a user's task. The MTEB maintainers caution that a predefined benchmark may contain only part of the tasks a user cares about, and the leaderboard flags models trained on benchmark data.[22][17]
  • Prompts and prefixes are part of the model. nomic-embed-text-v1.5 requires task prefixes such as search_document: and search_query:, multilingual-e5-large-instruct needs an instruction on the query side, and the Qwen3 card reports that instructions typically improve results by 1% to 5%.[29][39][6]
  • Language claims come with caveats. The multilingual E5 card warns of weaker results in low-resource languages, and Jina states that tuning of Jina Embeddings v3 focused on 30 of the 100 languages its base model covers.[32][37]
  • Short input limits truncate text. all-MiniLM-L6-v2 cuts input off at 256 word pieces and multilingual E5 at 512 tokens.[7][32]
  • Some weights are gated. The Hub marks the embeddinggemma-300m repository as gated and does not serve its files to anonymous requests.[11]

Figures in this article were checked on 19 November 2026. Leaderboard positions and download counts change over time.

Continue reading

More from See also

More to read

Topics:AI ModelsInformation RetrievalNatural Language ProcessingOpen Source AI

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25MTEB maintainers. "Benchmark Overview · MTEB Leaderboard". Benchmarks MTEB(Multilingual, v2), MTEB(eng, v2), MTEB(Code, v1), RTEB(beta) and ViDoRe(v3), read through the leaderboard's public API (mteb-leaderboard-backend.hf.space/docs) on 2026-11-19. leaderboard.mteb.org
  2. ^1 ^2 ^3 ^4 ^5 ^6 ^7Hugging Face. "Models" listing sorted by downloads, tasks Sentence Similarity, Feature Extraction and Text Ranking, read through the Hub API on 2026-11-19. huggingface.co/models
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Microsoft. "microsoft/harrier-oss-v1-27b" (model card). Hugging Face. huggingface.co/...harrier-oss-v1-27b
  4. ^1 ^2 ^3Tencent. "tencent/KaLM-Embedding-Gemma3-12B-2511" (model card). Hugging Face. huggingface.co/...KaLM-Embedding-Gemma3-12B-2511
  5. ^1 ^2 ^3 ^4NVIDIA. "nvidia/llama-embed-nemotron-8b" (model card). Hugging Face. huggingface.co/...llama-embed-nemotron-8b
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Qwen. "Qwen/Qwen3-Embedding-8B" (model card). Hugging Face. huggingface.co/...Qwen3-Embedding-8B
  7. ^1 ^2 ^3 ^4 ^5 ^6 ^7Sentence Transformers. "sentence-transformers/all-MiniLM-L6-v2" (model card). Hugging Face. huggingface.co/...all-MiniLM-L6-v2
  8. ^1 ^2Qwen Team. "Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models". Qwen blog, June 5, 2025. qwenlm.github.io/...qwen3-embedding
  9. ^1 ^2 ^3Nomic AI. "nomic-ai/nomic-embed-text-v2-moe" (model card). Hugging Face. huggingface.co/...nomic-embed-text-v2-moe
  10. ^1 ^2 ^3 ^4Nomic AI. "nomic-ai/colnomic-embed-multimodal-3b" (model card and LICENSE file). Hugging Face. huggingface.co/...colnomic-embed-multimodal-3b
  11. ^1 ^2 ^3Google. "google/embeddinggemma-300m" (gated model repository; metadata read through the Hub API on 2026-11-19). Hugging Face. huggingface.co/...embeddinggemma-300m
  12. ^1 ^2 ^3 ^4 ^5 ^6Google DeepMind. "EmbeddingGemma model card". Gemma Documentation. ai.google.dev/...model_card
  13. ^Hugging Face. "Models Download Stats". Hugging Face Hub documentation. huggingface.co/...models-download-stats
  14. ^Muennighoff, Niklas; Tazi, Nouamane; Magne, Loïc; Reimers, Nils. "MTEB: Massive Text Embedding Benchmark". arXiv:2210.07316, first version October 13, 2022, revised March 19, 2023 (v3). arxiv.org/...2210.07316
  15. ^1 ^2Enevoldsen, Kenneth; Chung, Isaac; Kerboua, Imene; Kardos, Márton; Mathur, Ashwin; et al. "MMTEB: Massive Multilingual Text Embedding Benchmark". arXiv:2502.13595, February 19, 2025. arxiv.org/...2502.13595
  16. ^1 ^2 ^3NVIDIA. "nvidia/NV-Embed-v2" (model card). Hugging Face. huggingface.co/...NV-Embed-v2
  17. ^1 ^2 ^3MTEB maintainers. "mteb/leaderboard/text_segments.py" (leaderboard FAQ text). GitHub, embeddings-benchmark/mteb, read 2026-11-19. github.com/...text_segments.py
  18. ^MTEB maintainers. "Run the Leaderboard". MTEB documentation. docs.mteb.org/...leaderboard
  19. ^MTEB maintainers. "Submit results". MTEB documentation source, GitHub, embeddings-benchmark/mteb. github.com/...submitting_results.md
  20. ^Liu, Frank; Enevoldsen, Kenneth; Solomatin, Roman; Chung, Isaac; Aarsen, Tom; Fődi, Zoltán. "Introducing RTEB: A New Standard for Retrieval Evaluation". Hugging Face Blog, October 1, 2025. huggingface.co/...rteb
  21. ^1 ^2 ^3 ^4 ^5 ^6 ^7Google DeepMind. "google/embeddinggemma-2" (model card). Hugging Face. huggingface.co/...embeddinggemma-2
  22. ^1 ^2Solomatin, Roman; Enevoldsen, Kenneth; Chung, Isaac. "MTEB Leaderboard: From a slow demo to feature-rich leaderboard". MTEB blog, June 12, 2026. github.com/...mteb-leaderboard.md
  23. ^1 ^2 ^3 ^4 ^5Octen. "Octen/Octen-Embedding-8B" (model card). Hugging Face. huggingface.co/...Octen-Embedding-8B
  24. ^1 ^2 ^3 ^4 ^5CodeFuse. "codefuse-ai/F2LLM-v2-14B" (model card). Hugging Face. huggingface.co/...F2LLM-v2-14B
  25. ^1 ^2 ^3 ^4 ^5 ^6 ^7BAAI. "BAAI/bge-small-en-v1.5" (model card). Hugging Face. huggingface.co/...bge-small-en-v1.5
  26. ^1 ^2 ^3 ^4 ^5 ^6BAAI. "BAAI/bge-m3" (model card). Hugging Face. huggingface.co/...bge-m3
  27. ^1 ^2 ^3Sentence Transformers. "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2" (model card). Hugging Face. huggingface.co/...hrase-multilingual-MiniLM-L12-v2
  28. ^1 ^2Sentence Transformers. "sentence-transformers/all-mpnet-base-v2" (model card). Hugging Face. huggingface.co/...all-mpnet-base-v2
  29. ^1 ^2 ^3 ^4 ^5Nomic AI. "nomic-ai/nomic-embed-text-v1.5" (model card). Hugging Face. huggingface.co/...nomic-embed-text-v1.5
  30. ^1 ^2intfloat. "intfloat/multilingual-e5-small" (model card). Hugging Face. huggingface.co/...multilingual-e5-small
  31. ^1 ^2 ^3Sentence Transformers. "sentence-transformers/paraphrase-multilingual-mpnet-base-v2" (model card). Hugging Face. huggingface.co/...hrase-multilingual-mpnet-base-v2
  32. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8intfloat. "intfloat/multilingual-e5-large" (model card). Hugging Face. huggingface.co/...multilingual-e5-large
  33. ^1 ^2 ^3IBM. "ibm-granite/granite-embedding-small-english-r2" (model card). Hugging Face. huggingface.co/...granite-embedding-small-english-r2
  34. ^1 ^2intfloat. "intfloat/multilingual-e5-base" (model card). Hugging Face. huggingface.co/...multilingual-e5-base
  35. ^1 ^2 ^3 ^4Snowflake. "Snowflake/snowflake-arctic-embed-l-v2.0" (model card). Hugging Face. huggingface.co/...snowflake-arctic-embed-l-v2.0
  36. ^1 ^2 ^3 ^4Alibaba DAMO Academy. "thenlper/gte-small" (model card). Hugging Face. huggingface.co/...gte-small
  37. ^1 ^2 ^3 ^4Jina AI. "jinaai/jina-embeddings-v3" (model card). Hugging Face. huggingface.co/...jina-embeddings-v3
  38. ^1 ^2 ^3 ^4Dua, Sahil; Schechter Vera, Henrique. "EmbeddingGemma 2: an open, lightweight multimodal embedding model". Google, The Keyword, October 6, 2026. blog.google/...embeddinggemma-2
  39. ^1 ^2intfloat. "intfloat/multilingual-e5-large-instruct" (model card). Hugging Face. huggingface.co/...multilingual-e5-large-instruct
  40. ^1 ^2 ^3 ^4Alibaba. "Alibaba-NLP/gte-multilingual-base" (model card). Hugging Face. huggingface.co/...gte-multilingual-base
  41. ^Alibaba. "Alibaba-NLP/gte-modernbert-base" (model card). Hugging Face. huggingface.co/...gte-modernbert-base
  42. ^1 ^2Mixedbread. "mixedbread-ai/mxbai-embed-large-v1" (model card). Hugging Face. huggingface.co/...mxbai-embed-large-v1
  43. ^1 ^2 ^3 ^4NovaSearch. "NovaSearch/stella_en_400M_v5" (model card). Hugging Face. huggingface.co/...stella_en_400M_v5
  44. ^1 ^2 ^3IBM. "ibm-granite/granite-embedding-english-r2" (model card). Hugging Face. huggingface.co/...granite-embedding-english-r2
  45. ^1 ^2 ^3IBM. "ibm-granite/granite-embedding-311m-multilingual-r2" (model card). Hugging Face. huggingface.co/...e-embedding-311m-multilingual-r2
  46. ^1 ^2Voyage AI. "MongoDB/voyage-4-nano" (model card). Hugging Face. huggingface.co/...voyage-4-nano
  47. ^1 ^2Perplexity. "perplexity-ai/pplx-embed-v1-0.6b" (model card). Hugging Face. huggingface.co/...pplx-embed-v1-0.6b
  48. ^1 ^2NVIDIA. "nvidia/Nemotron-3-Embed-8B-BF16" (model card). Hugging Face. huggingface.co/...Nemotron-3-Embed-8B-BF16
  49. ^OpenMDW. "OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)". openmdw.ai. openmdw.ai/...1-1
  50. ^Google. "Gemma Terms of Use". Gemma Documentation, last modified April 1, 2026. ai.google.dev/...terms
  51. ^BAAI. "BAAI/bge-multilingual-gemma2" (model card). Hugging Face. huggingface.co/...bge-multilingual-gemma2
  52. ^1 ^2 ^3NVIDIA. "nvidia/llama-nemotron-embed-1b-v2" (model card). Hugging Face. huggingface.co/...llama-nemotron-embed-1b-v2
  53. ^1 ^2 ^3NVIDIA. "nvidia/llama-nemotron-rerank-1b-v2" (model card and commit history). Hugging Face. huggingface.co/...llama-nemotron-rerank-1b-v2
  54. ^1 ^2Tencent. "LICENSE.txt", tencent/KaLM-Embedding-Gemma3-12B-2511. Hugging Face. huggingface.co/...LICENSE.txt
  55. ^1 ^2 ^3ViDoRe. "vidore/colpali-v1.3" (model card). Hugging Face. huggingface.co/...colpali-v1.3
  56. ^Jina AI. "jinaai/jina-embeddings-v5-text-small" (model card). Hugging Face. huggingface.co/...jina-embeddings-v5-text-small
  57. ^1 ^2 ^3 ^4 ^5Jina AI. "jinaai/jina-embeddings-v4" (model card, LICENSE file and commit history). Hugging Face. huggingface.co/...jina-embeddings-v4
  58. ^1 ^2Jina AI. "jinaai/jina-reranker-v3" (model card). Hugging Face. huggingface.co/...jina-reranker-v3
  59. ^1 ^2Jina AI. "jinaai/jina-colbert-v2" (model card). Hugging Face. huggingface.co/...jina-colbert-v2
  60. ^Jina AI. "jinaai/jina-clip-v2" (model card). Hugging Face. huggingface.co/...jina-clip-v2
  61. ^1 ^2Jina AI. "jinaai/jina-code-embeddings-1.5b" (model card). Hugging Face. huggingface.co/...jina-code-embeddings-1.5b
  62. ^Salesforce. "Salesforce/SFR-Embedding-Mistral" (model card). Hugging Face. huggingface.co/...SFR-Embedding-Mistral
  63. ^Linq AI Research. "Linq-AI-Research/Linq-Embed-Mistral" (model card). Hugging Face. huggingface.co/...Linq-Embed-Mistral
  64. ^1 ^2 ^3ViDoRe. "vidore/colqwen2.5-v0.2" (model card). Hugging Face. huggingface.co/...colqwen2.5-v0.2
  65. ^1 ^2LightOn. "lightonai/Reason-ModernColBERT" (model card). Hugging Face. huggingface.co/...Reason-ModernColBERT
  66. ^1 ^2ZeroEntropy. "zeroentropy/zerank-2-reranker" (model card and commit history). Hugging Face. huggingface.co/...zerank-2-reranker
  67. ^Choi, Min; Dua, Sahil; Lisak, Alice. "Introducing EmbeddingGemma: The Best-in-Class Open Model for On-Device Embeddings". Google Developers Blog, September 4, 2025. developers.googleblog.com/...ducing-embeddinggemma
  68. ^Nomic AI. "nomic-ai/nomic-embed-code" (model card). Hugging Face. huggingface.co/...nomic-embed-code
  69. ^Qwen. "Qwen/Qwen3-VL-Embedding-8B" (model card). Hugging Face. huggingface.co/...Qwen3-VL-Embedding-8B
  70. ^Tencent. "tencent/WeMM-Embedding-9B" (model card and LICENSE file). Hugging Face. huggingface.co/...WeMM-Embedding-9B
  71. ^Google. "google/siglip2-so400m-patch14-384" (model card). Hugging Face. huggingface.co/...siglip2-so400m-patch14-384
  72. ^OpenAI. "openai/clip-vit-large-patch14" (model card). Hugging Face. huggingface.co/...clip-vit-large-patch14
  73. ^Faysse, Manuel; Sibille, Hugues; Wu, Tony; Omrani, Bilel; Viaud, Gautier; Hudelot, Céline; Colombo, Pierre. "ColPali: Efficient Document Retrieval with Vision Language Models". arXiv:2407.01449, June 27, 2024. arxiv.org/...2407.01449
  74. ^Nomic AI. "nomic-ai/colnomic-embed-multimodal-7b" (model card). Hugging Face. huggingface.co/...colnomic-embed-multimodal-7b
  75. ^ViDoRe. "vidore/colqwen2-v1.0" (model card). Hugging Face. huggingface.co/...colqwen2-v1.0
  76. ^Khattab, Omar; Zaharia, Matei. "ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT". arXiv:2004.12832, April 27, 2020. arxiv.org/...2004.12832
  77. ^1 ^2 ^3ColBERT. "colbert-ir/colbertv2.0" (model card). Hugging Face. huggingface.co/...colbertv2.0
  78. ^LightOn. "lightonai/GTE-ModernColBERT-v1" (model card). Hugging Face. huggingface.co/...GTE-ModernColBERT-v1
  79. ^Answer.AI. "answerdotai/answerai-colbert-small-v1" (model card). Hugging Face. huggingface.co/...answerai-colbert-small-v1
  80. ^Mixedbread. "mixedbread-ai/mxbai-edge-colbert-v0-32m" (model card). Hugging Face. huggingface.co/...mxbai-edge-colbert-v0-32m
  81. ^1 ^2Sentence Transformers. "SentenceTransformers Documentation". sbert.net. sbert.net
  82. ^1 ^2Sentence Transformers. "cross-encoder/ms-marco-MiniLM-L6-v2" (model card). Hugging Face. huggingface.co/...ms-marco-MiniLM-L6-v2
  83. ^BAAI. "BAAI/bge-reranker-v2-m3" (model card). Hugging Face. huggingface.co/...bge-reranker-v2-m3
  84. ^Qwen. "Qwen/Qwen3-Reranker-8B" (model card). Hugging Face. huggingface.co/...Qwen3-Reranker-8B
  85. ^Alibaba. "Alibaba-NLP/gte-reranker-modernbert-base" (model card). Hugging Face. huggingface.co/...gte-reranker-modernbert-base
  86. ^Mixedbread. "mixedbread-ai/mxbai-rerank-large-v2" (model card). Hugging Face. huggingface.co/...mxbai-rerank-large-v2
  87. ^Jina AI. "jinaai/jina-reranker-v2-base-multilingual" (model card). Hugging Face. huggingface.co/...jina-reranker-v2-base-multilingual
  88. ^Kusupati, Aditya; Bhatt, Gantavya; Rege, Aniket; Wallingford, Matthew; Sinha, Aditya; et al. "Matryoshka Representation Learning". arXiv:2205.13147, May 26, 2022. arxiv.org/...2205.13147
  89. ^Microsoft. "MS MARCO". microsoft.github.io/msmarco
  90. ^Hugging Face. "huggingface/sentence-transformers". GitHub repository, figures as of November 2026. github.com/...sentence-transformers
  91. ^1 ^2 ^3Hugging Face. "huggingface/text-embeddings-inference". GitHub repository (README, LICENSE and commit history), figures as of November 2026. github.com/...text-embeddings-inference
  92. ^1 ^2Ollama. "Embeddings". Ollama documentation. docs.ollama.com/...embeddings
  93. ^Ollama. "ollama/ollama". GitHub repository, figures as of November 2026. github.com/...ollama
  94. ^1 ^2 ^3ggml-org. "LLaMA.cpp HTTP Server" (tools/server/README.md). GitHub, ggml-org/llama.cpp. github.com/...README.md
  95. ^ggml-org. "ggml-org/llama.cpp". GitHub repository, figures as of November 2026. github.com/...llama.cpp
  96. ^1 ^2vLLM project. "Pooling Models". vLLM documentation (latest developer preview). docs.vllm.ai/...pooling_models
  97. ^vLLM project. "vllm-project/vllm". GitHub repository, figures as of November 2026. github.com/...vllm
  98. ^michaelfeil. "michaelfeil/infinity". GitHub repository (README and LICENSE), figures as of November 2026. github.com/...infinity
  99. ^Qdrant. "qdrant/fastembed". GitHub repository (README and LICENSE), figures as of November 2026. github.com/...fastembed
  100. ^Hugging Face. "HfApi Client" (the created_at field of ModelInfo). huggingface_hub documentation. huggingface.co/...hf_api
  101. ^Jina AI. "Jina Embeddings v3: A Frontier Multilingual Embedding Model". Jina AI News, September 18, 2024. jina.ai/...a-frontier-multilingual-embedding-model
  102. ^OpenAI. "openai/CLIP". GitHub repository (LICENSE). github.com/...CLIP

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 7,609 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check before publication on November 19, 2026: 340 claims checked against 236 cited and independent sources; 2 material and 21 minor corrections applied before publication.

Cite this page: AI Wiki. "Best open-weight embedding models." aiwiki.ai, updated 19 Nov 2026, fact-checked 19 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/best_open_weight_embedding_models

Edit article