Best open-weight embedding models
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.
| Test | Criterion | Source of the figure |
|---|---|---|
| Leaderboard | One 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] |
| Downloads | A 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 listings | Hugging Face Hub API, read 2026-11-19 [2] |
| Illustration | A 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 use | The model's own card |
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) | Tasks | What it measures, per the leaderboard | Version note |
|---|---|---|---|
| MTEB(Multilingual, v2) | 131 | Text embedding quality across 250+ languages: bitext mining, classification, clustering, retrieval, reranking, semantic similarity | Replaced v1 (132 tasks) after one dataset was removed from the Hugging Face Hub |
| MTEB(eng, v2) | 41 | English classification, clustering, pair classification, reranking, retrieval, semantic similarity, summarization | Replaced v1 (56 tasks) |
| MTEB(Code, v1) | 12 | Code retrieval: code search, text-to-SQL, code feedback | |
| RTEB(beta) | 30 | Retrieval in legal, finance, code and healthcare domains, on open and closed datasets | Beta |
| ViDoRe(v3) | 10 | Retrieval over images of enterprise documents, on open and closed datasets |
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) | Model | Developer | Weights licence | Parameters | Dimensions | Max input tokens | Zero-shot | Mean(Task) | Mean(TaskType) |
|---|---|---|---|---|---|---|---|---|---|
| 1 | harrier-oss-v1-27b | Microsoft | MIT [3] | 27.0B | 5,376 | 32,768 | 78% | 74.27 | 64.20 |
| 2 | KaLM-Embedding-Gemma3-12B-2511 | Tencent | Tencent KaLM licence (not intended for use in the EU) plus Gemma terms [4] | 11.8B | 3,840 | 32,768 | 73% | 72.32 | 62.51 |
| 3 | llama-embed-nemotron-8b | NVIDIA | NVIDIA License (non-commercial) plus Llama 3.1 licence [5] | 7.5B | 4,096 | 32,768 | 99% | 69.46 | 61.09 |
| 4 | Qwen3-Embedding-8B | Alibaba (Qwen) | Apache 2.0 [6] | 7.6B | 4,096 | 32,768 | 99% | 70.58 | 61.69 |
| 6 | Qwen3-Embedding-4B | Alibaba (Qwen) | Apache 2.0 [6] | 4.0B | 2,560 | 32,768 | 99% | 69.45 | 60.86 |
| 7 | Octen-Embedding-8B | Octen | Apache 2.0 [23] | 7.6B | 4,096 | 32,768 | 99% | 67.84 | 60.28 |
| 8 | F2LLM-v2-14B | CodeFuse | Apache 2.0 [24] | 14.0B | 5,120 | 40,960 | 88% | 68.74 | 59.45 |
| 9 | F2LLM-v2-8B | CodeFuse | Apache 2.0 [24] | 7.6B | 4,096 | 40,960 | 88% | 68.09 | 58.99 |
| 10 | harrier-oss-v1-0.6b | Microsoft | MIT [3] | 0.6B | 1,024 | 32,768 | 78% | 69.01 | 59.00 |
| 12 | F2LLM-v2-4B | CodeFuse | Apache 2.0 [24] | 4.0B | 2,560 | 40,960 | 88% | 67.06 | 58.25 |
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]
| Model | Developer | Weights licence | Downloads, previous month (millions) | Parameters | Dimensions | Max input | MTEB(eng, v2) | MTEB(Multilingual, v2) |
|---|---|---|---|---|---|---|---|---|
| all-MiniLM-L6-v2 | Sentence Transformers | Apache 2.0 [7] | 169.03 | 22.7M | 384 | 256 word pieces | 151st, 59.03 | 152nd, 41.39 |
| bge-small-en-v1.5 | BAAI | MIT [25] | 43.96 | 33.4M | 384 | 512 tokens | 89th, 64.30 | 122nd, 43.76 |
| bge-m3 | BAAI | MIT [26] | 31.74 | 568M | 1,024 | 8,192 tokens | 225th, no mean | 52nd, no mean |
| paraphrase-multilingual-MiniLM-L12-v2 | Sentence Transformers | Apache 2.0 [27] | 25.22 | 117.7M | 384 | 128 (card); 512 (leaderboard) | 195th, 56.42 | 104th, 48.73 |
| all-mpnet-base-v2 | Sentence Transformers | Apache 2.0 [28] | 19.90 | 109.5M | 768 | 384 word pieces | 128th, no mean | 139th, no mean |
| nomic-embed-text-v1.5 | Nomic AI | Apache 2.0 [29] | 11.52 | 136.7M | 768 | 8,192 tokens | 123rd, 62.20 | 117th, 44.10 |
| multilingual-e5-small | intfloat | MIT [30] | 9.45 | 117.7M | 384 | 512 tokens | 161st, 59.69 | 73rd, no mean |
| paraphrase-multilingual-mpnet-base-v2 | Sentence Transformers | Apache 2.0 [31] | 9.38 | 278.0M | 768 | 128 (card); 512 (leaderboard) | 158th, 58.27 | 83rd, 51.93 |
| bge-base-en-v1.5 | BAAI | MIT [25] | 8.60 | 109.5M | 768 | 512 tokens | 77th, 65.14 | 111th, 44.52 |
| Qwen3-Embedding-0.6B | Alibaba (Qwen) | Apache 2.0 [6] | 7.89 | 595.8M | 1,024 | 32,768 tokens | 28th, 70.47 | 20th, 64.34 |
| multilingual-e5-large | intfloat | MIT [32] | 7.33 | 559.9M | 1,024 | 512 tokens | 128th, no mean | 54th, no mean |
| bge-large-en-v1.5 | BAAI | MIT [25] | 7.05 | 335.1M | 1,024 | 512 tokens | 69th, 65.89 | 98th, 45.08 |
| granite-embedding-small-english-r2 | IBM | Apache 2.0 [33] | 6.66 | 47.7M | 384 | 8,192 tokens | 132nd, 61.07 | 243rd, no mean |
| multilingual-e5-base | intfloat | MIT [34] | 4.70 | 278.0M | 768 | 512 tokens | 146th, 61.00 | 69th, 57.02 |
| snowflake-arctic-embed-l-v2.0 | Snowflake | Apache 2.0 [35] | 4.64 | 567.8M | 1,024 | 8,192 tokens | 104th, 63.60 | 64th, 57.03 |
| gte-small | Alibaba DAMO Academy | MIT [36] | 4.06 | 33.4M | 384 | 512 tokens | 96th, 63.22 | 123rd, no mean |
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 line | Developer | Weights licence | Released | Parameters | Dimensions | Max input tokens | Languages, per the card |
|---|---|---|---|---|---|---|---|
| Qwen3 Embedding 0.6B, 4B, 8B | Alibaba (Qwen) | Apache 2.0 [6][8] | 2025-06-05 | 0.6B, 4B, 8B | 1,024, 2,560, 4,096 | 32k | Over 100, plus programming languages |
| harrier-oss-v1 270m, 0.6b, 27b | Microsoft | MIT [3] | repo 2026-03-30 | 270M, 0.6B, 27B | 640, 1,024, 5,376 | 32,768 | Multilingual |
| EmbeddingGemma 2 | Google DeepMind | Apache 2.0 [21][38] | 2026-10-06 | 740M (270M text-only) | 768 | 8,192 | Over 100 |
| BGE-M3 | BAAI | MIT [26] | 2024-01-30 [25] | 568M | 1,024 | 8,192 | Over 100 |
| BGE v1.5 English: small, base, large | BAAI | MIT [25] | repo 2023-09 | 33.4M, 109.5M, 335.1M | 384, 768, 1,024 | 512 | English |
| Multilingual E5: small, base, large, large-instruct | intfloat | MIT [32][39] | repo 2023-05 to 2024-02 | 117.7M, 278.0M, 559.9M, 559.9M | 384, 768, 1,024, 1,024 | 512 | 100, per the card of the large model |
| gte-multilingual-base | Alibaba | Apache 2.0 [40] | repo 2024-07-20 | 305M | 768 | 8,192 | Over 70 |
| gte-modernbert-base | Alibaba (Tongyi Lab) | Apache 2.0 [41] | repo 2025-01-20 | 149M | 768 | 8,192 | English |
| nomic-embed-text-v1.5 | Nomic AI | Apache 2.0 [29] | repo 2024-02-10 | 136.7M | 768 | 8,192 | English |
| nomic-embed-text-v2-moe | Nomic AI | Apache 2.0 [9] | repo 2025-02-07 | 475M (305M active) | 768 | 512 | About 100 |
| snowflake-arctic-embed-l-v2.0 | Snowflake | Apache 2.0 [35] | 2024-12-04 | 568M | 1,024 | 8,192 | Multilingual |
| mxbai-embed-large-v1 | Mixedbread | Apache 2.0 [42] | repo 2024-03-07 | 335.1M | 1,024 | 512 | English |
| Stella v5: stella_en_400M_v5, stella_en_1.5B_v5 | NovaSearch | MIT [43] | repo 2024-07-12 | 435M, 1.54B | 512 to 8,192 (default 1,024) | 8,192 on the leaderboard (400M model); the card recommends 512 | English |
| Granite Embedding R2: small English, English, 311M multilingual | IBM | Apache 2.0 [33][44][45] | 2025-08-15 (English), 2026-04-29 (multilingual) | 47M, 149M, 311M | 384, 768, 768 | 8,192, 8,192, 32,768 | English; over 200 for the multilingual model |
| all-MiniLM-L6-v2, all-mpnet-base-v2 | Sentence Transformers | Apache 2.0 [7][28] | first commit 2021-08-30 | 22.7M, 109.5M | 384, 768 | 256 and 384 word pieces | English |
| paraphrase-multilingual-MiniLM-L12-v2, paraphrase-multilingual-mpnet-base-v2 | Sentence Transformers | Apache 2.0 [27][31] | first commit 2021-06-02 | 117.7M, 278.0M | 384, 768 | 128 on the cards; 512 on the leaderboard | Multilingual |
| gte-small | Alibaba DAMO Academy | MIT [36] | repo 2023-07-27 | 33.4M | 384 | 512 | English |
| voyage-4-nano | Voyage AI (MongoDB) | Apache 2.0 [46] | repo 2026-01-06 | 346M | 2,048, 1,024, 512 or 256 | 32,000 | Multilingual |
| pplx-embed-v1 0.6b, 4b | Perplexity | MIT [47] | repo 2026-01-14 (0.6b), 2026-01-20 (4b) | 0.6B, 4B | 1,024, 2,560 | 32K | Multilingual |
| Octen-Embedding-8B | Octen | Apache 2.0 [23] | repo 2025-12-23 | 7.6B | 4,096 | 32,768 | Over 100 |
| F2LLM-v2 (eight sizes) | CodeFuse | Apache 2.0 [24] | repo 2026-03 | 80M to 14B | Up to 5,120 | 40,960 | Over 200 |
| Nemotron-3-Embed-8B | NVIDIA | OpenMDW 1.1 [48] | 2026-07-16 | 7.95B | 4,096 | 32,768 | Multilingual |
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
| Model | Developer | Terms | Conditions stated |
|---|---|---|---|
| EmbeddingGemma (embeddinggemma-300m) | Google DeepMind | Gemma 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-gemma2 | BAAI | Gemma licence tag on the Hub [51] | Gemma terms apply |
| llama-nemotron-embed-1b-v2 | NVIDIA | NVIDIA 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-v2 | NVIDIA | OpenMDW 1.1 plus the Llama 3.2 community licence [53] | Two licences apply together |
| KaLM-Embedding-Gemma3-12B-2511 | Tencent | Tencent KaLM licence plus Gemma terms for the base model [4][54] | "KaLM-Embedding IS NOT INTENDED FOR USE WITHIN THE EUROPEAN UNION." |
| ColPali v1.3 | ViDoRe (ColPali authors) | Gemma licence on the PaliGemma backbone, MIT on the adapters [55] | The Hub tag reads MIT, which covers the adapters only |
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
| Model | Developer | Terms | What the card or licence says |
|---|---|---|---|
| Jina Embeddings v3 | Jina AI | CC 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-small | Jina AI | CC BY-NC 4.0 [56] | "For commercial use, please contact us" |
| Jina Embeddings v4 | Jina AI | Qwen Research License [57] | Licence grants use for non-commercial purposes only |
| jina-reranker-v3, jina-colbert-v2, jina-clip-v2, jina-code-embeddings | Jina AI | CC 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-v2 | NVIDIA | CC BY-NC 4.0 [16] | "This model should not be used for any commercial purpose." |
| llama-embed-nemotron-8b | NVIDIA | NVIDIA License (non-commercial) plus the Llama 3.1 community licence [5] | "This model is for non-commercial/research use only" |
| SFR-Embedding-Mistral | Salesforce | CC BY-NC 4.0 [62] | Research purposes only |
| Linq-Embed-Mistral | Linq AI Research | CC BY-NC 4.0 [63] | Research purposes only |
| colnomic-embed-multimodal-3b | Nomic AI | Qwen Research License file, no Hub licence tag [10] | "Model weights available for research use" |
| ColQwen2.5 v0.2 | ViDoRe (ColPali authors) | Qwen Research License on the backbone, MIT on the adapters [64] | The Hub tag reads MIT, which covers the adapters only |
| Reason-ModernColBERT | LightOn | CC BY-NC 4.0 [65] | Non-commercial because of its training data |
For commercial use NVIDIA's NV-Embed-v2 card recommends the models in its NeMo Retriever microservices (NIMs) instead.[16]
Licence changes
| Date | Model | Change | Evidence |
|---|---|---|---|
| 2025-08-19 | jina-embeddings-v4 | Licence 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 3B | Card and repository history [57] |
| 2026-07-24 | zerank-1, zerank-2 (ZeroEntropy rerankers) | CC BY-NC 4.0 to Apache 2.0 | Repository history: commits titled "Relicense model under Apache 2.0" [66] |
| 2026-08-26 | llama-nemotron-rerank-1b-v2 | NVIDIA Open Model License to OpenMDW 1.1; the Llama 3.2 community licence still applies | Repository history: commit titled "Update license to OpenMDW-1.1" [53] |
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
| Model | Language coverage stated by the developer | Licence |
|---|---|---|
| BGE-M3 | More than 100 working languages [26] | MIT |
| Multilingual E5 | The 100 languages of XLM-RoBERTa, with the caveat that "low-resource languages may see performance degradation" [32] | MIT |
| Qwen3 Embedding | Over 100 languages [6] | Apache 2.0 |
| gte-multilingual-base | Over 70 languages [40] | Apache 2.0 |
| EmbeddingGemma, EmbeddingGemma 2 | Over 100 languages [12][21] | Gemma terms; Apache 2.0 |
| granite-embedding-311m-multilingual-r2 | Over 200 languages, with enhanced support for 52 [45] | Apache 2.0 |
| nomic-embed-text-v2-moe | About 100 languages [9] | Apache 2.0 |
| F2LLM-v2 | More than 200 languages [24] | Apache 2.0 |
| Jina Embeddings v3 | The 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 |
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]
| Reranker | Developer | Weights licence | Parameters | Downloads, previous month | Notes from the card |
|---|---|---|---|---|---|
| cross-encoder/ms-marco-MiniLM-L6-v2 | Sentence Transformers | Apache 2.0 [82] | 22.7M | 56.93M | Trained on MS MARCO passage ranking |
| bge-reranker-v2-m3 | BAAI | Apache 2.0 [83] | 567.8M | 14.83M | Multilingual |
| Qwen3-Reranker 0.6B, 4B, 8B | Alibaba (Qwen) | Apache 2.0 [84] | 0.6B, 4B, 8B | 1.16M, 2.87M, 0.06M | 32k context, over 100 languages |
| gte-reranker-modernbert-base | Alibaba | Apache 2.0 [85] | 149.6M | 1.38M | 8,192 tokens |
| mxbai-rerank-base-v2, mxbai-rerank-large-v2 | Mixedbread | Apache 2.0 [86] | 494M, 1.54B | 0.02M, 0.01M | Over 100 languages |
| zerank-2 | ZeroEntropy | Apache 2.0 since 2026-07-24 [66] | 4.0B | 0.05M | Previously non-commercial |
| llama-nemotron-rerank-1b-v2 | NVIDIA | OpenMDW 1.1 plus Llama 3.2 licence [53] | 1.24B | 0.06M | See licence changes |
| jina-reranker-v2-base-multilingual | Jina AI | CC BY-NC 4.0 [87] | 278.4M | 1.11M | |
| jina-reranker-v3 | Jina AI | CC BY-NC 4.0 [58] | 596.8M | 0.74M | Listwise, up to 64 documents per call |
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.
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]
| Family | Size | Dimensions | Rank (Borda) | Mean(Task) |
|---|---|---|---|---|
| harrier-oss-v1 | 270M | 640 | 19 | 66.55 |
| harrier-oss-v1 | 0.6B | 1,024 | 10 | 69.01 |
| harrier-oss-v1 | 27B | 5,376 | 1 | 74.27 |
| Qwen3 Embedding | 0.6B | 1,024 | 20 | 64.34 |
| Qwen3 Embedding | 4B | 2,560 | 6 | 69.45 |
| Qwen3 Embedding | 8B | 4,096 | 4 | 70.58 |
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]
| Model | What 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] |
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.
| Tool | Maintainer | Licence | Latest release | Stars | What its documentation says |
|---|---|---|---|---|---|
| Sentence Transformers | Hugging Face (created by UKP Lab) | Apache 2.0 | v6.3.0, 2026-11-11 | 19,289 | Python library for using and training embedding, reranker, sparse and multi-vector models [81][90] |
| Text Embeddings Inference (TEI) | Hugging Face | Apache 2.0 | v1.9.4, 2026-09-15 | 5,128 | Server for embedding and sequence-classification (reranking) models, with CPU and CUDA Docker images, Metal support on Macs and an OpenAI-compatible route [91] |
| Ollama | Ollama | MIT | v0.40.16, 2026-11-09 | 187,552 | /api/embed endpoint that returns L2-normalized vectors [92][93] |
| llama.cpp | ggml-org | MIT | v1.0.0, 2026-11-13 | 135,610 | Server flags --embedding and --reranking, with /v1/embeddings and /v1/rerank endpoints [94][95] |
| vLLM | vLLM project | Apache 2.0 | v0.34.0, 2026-11-16 | 97,184 | Pooling models for embedding, classification and scoring [96][97] |
| Infinity | michaelfeil (GitHub account) | MIT | v3.0.0, 2026-10-26 | 2,987 | REST server for embeddings, reranking and classification; v3 is written in Rust on LibTorch [98] |
| FastEmbed | Qdrant | Apache 2.0 | v0.9.1, 2026-10-24 | 3,303 | Python library on ONNX Runtime with no PyTorch dependency; dense, sparse, late-interaction and image embeddings, plus rerankers [99] |
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:andsearch_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
- List of open-weight modelsThis is a list of open-weight models: artificial intelligence models whose trained parameters (weights) the developer has published for download, so that the model can be run, studied and in most cases…Listed in See also
- Timeline of open-weight modelsA timeline of open-weight models lists, in date order, the releases of machine-learning models whose trained parameters (the weights) can be downloaded, together with the licence and policy events that changed…Listed in See also
- Best open-weight image generation modelsBest open-weight image generation models is a comparison of image generation and image editing models whose trained weights can be downloaded and run on a user's own hardware, current to 19 November 2026.Listed in See also
- Best Local and On-Device LLMsAs of November 2026, the best local LLM to run offline depends almost entirely on how much memory you can give it, so this guide ranks picks by hardware tier.Listed in See also
- How to Run LLMs LocallyThis guide explains how to run LLMs locally: downloading the weights of a large language model (LLM) and performing inference on hardware you control, instead of sending prompts to a hosted API.Listed in See also
- LocalAI (software)LocalAI is an open-source server for running AI models on hardware the operator controls. It is written mainly in Go and published under the MIT License.Listed in See also
More from See also
More to read
Topics:AI ModelsInformation RetrievalNatural Language ProcessingOpen Source AI
References
- ^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
- ^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
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Microsoft. "microsoft/harrier-oss-v1-27b" (model card). Hugging Face. huggingface.co/...harrier-oss-v1-27b
- ^1 ^2 ^3Tencent. "tencent/KaLM-Embedding-Gemma3-12B-2511" (model card). Hugging Face. huggingface.co/...KaLM-Embedding-Gemma3-12B-2511
- ^1 ^2 ^3 ^4NVIDIA. "nvidia/llama-embed-nemotron-8b" (model card). Hugging Face. huggingface.co/...llama-embed-nemotron-8b
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Qwen. "Qwen/Qwen3-Embedding-8B" (model card). Hugging Face. huggingface.co/...Qwen3-Embedding-8B
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Sentence Transformers. "sentence-transformers/all-MiniLM-L6-v2" (model card). Hugging Face. huggingface.co/...all-MiniLM-L6-v2
- ^1 ^2Qwen Team. "Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models". Qwen blog, June 5, 2025. qwenlm.github.io/...qwen3-embedding
- ^1 ^2 ^3Nomic AI. "nomic-ai/nomic-embed-text-v2-moe" (model card). Hugging Face. huggingface.co/...nomic-embed-text-v2-moe
- ^1 ^2 ^3 ^4Nomic AI. "nomic-ai/colnomic-embed-multimodal-3b" (model card and LICENSE file). Hugging Face. huggingface.co/...colnomic-embed-multimodal-3b
- ^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
- ^1 ^2 ^3 ^4 ^5 ^6Google DeepMind. "EmbeddingGemma model card". Gemma Documentation. ai.google.dev/...model_card
- ^Hugging Face. "Models Download Stats". Hugging Face Hub documentation. huggingface.co/...models-download-stats
- ^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
- ^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
- ^1 ^2 ^3NVIDIA. "nvidia/NV-Embed-v2" (model card). Hugging Face. huggingface.co/...NV-Embed-v2
- ^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
- ^MTEB maintainers. "Run the Leaderboard". MTEB documentation. docs.mteb.org/...leaderboard
- ^MTEB maintainers. "Submit results". MTEB documentation source, GitHub, embeddings-benchmark/mteb. github.com/...submitting_results.md
- ^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
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Google DeepMind. "google/embeddinggemma-2" (model card). Hugging Face. huggingface.co/...embeddinggemma-2
- ^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
- ^1 ^2 ^3 ^4 ^5Octen. "Octen/Octen-Embedding-8B" (model card). Hugging Face. huggingface.co/...Octen-Embedding-8B
- ^1 ^2 ^3 ^4 ^5CodeFuse. "codefuse-ai/F2LLM-v2-14B" (model card). Hugging Face. huggingface.co/...F2LLM-v2-14B
- ^1 ^2 ^3 ^4 ^5 ^6 ^7BAAI. "BAAI/bge-small-en-v1.5" (model card). Hugging Face. huggingface.co/...bge-small-en-v1.5
- ^1 ^2 ^3 ^4 ^5 ^6BAAI. "BAAI/bge-m3" (model card). Hugging Face. huggingface.co/...bge-m3
- ^1 ^2 ^3Sentence Transformers. "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2" (model card). Hugging Face. huggingface.co/...hrase-multilingual-MiniLM-L12-v2
- ^1 ^2Sentence Transformers. "sentence-transformers/all-mpnet-base-v2" (model card). Hugging Face. huggingface.co/...all-mpnet-base-v2
- ^1 ^2 ^3 ^4 ^5Nomic AI. "nomic-ai/nomic-embed-text-v1.5" (model card). Hugging Face. huggingface.co/...nomic-embed-text-v1.5
- ^1 ^2intfloat. "intfloat/multilingual-e5-small" (model card). Hugging Face. huggingface.co/...multilingual-e5-small
- ^1 ^2 ^3Sentence Transformers. "sentence-transformers/paraphrase-multilingual-mpnet-base-v2" (model card). Hugging Face. huggingface.co/...hrase-multilingual-mpnet-base-v2
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8intfloat. "intfloat/multilingual-e5-large" (model card). Hugging Face. huggingface.co/...multilingual-e5-large
- ^1 ^2 ^3IBM. "ibm-granite/granite-embedding-small-english-r2" (model card). Hugging Face. huggingface.co/...granite-embedding-small-english-r2
- ^1 ^2intfloat. "intfloat/multilingual-e5-base" (model card). Hugging Face. huggingface.co/...multilingual-e5-base
- ^1 ^2 ^3 ^4Snowflake. "Snowflake/snowflake-arctic-embed-l-v2.0" (model card). Hugging Face. huggingface.co/...snowflake-arctic-embed-l-v2.0
- ^1 ^2 ^3 ^4Alibaba DAMO Academy. "thenlper/gte-small" (model card). Hugging Face. huggingface.co/...gte-small
- ^1 ^2 ^3 ^4Jina AI. "jinaai/jina-embeddings-v3" (model card). Hugging Face. huggingface.co/...jina-embeddings-v3
- ^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
- ^1 ^2intfloat. "intfloat/multilingual-e5-large-instruct" (model card). Hugging Face. huggingface.co/...multilingual-e5-large-instruct
- ^1 ^2 ^3 ^4Alibaba. "Alibaba-NLP/gte-multilingual-base" (model card). Hugging Face. huggingface.co/...gte-multilingual-base
- ^Alibaba. "Alibaba-NLP/gte-modernbert-base" (model card). Hugging Face. huggingface.co/...gte-modernbert-base
- ^1 ^2Mixedbread. "mixedbread-ai/mxbai-embed-large-v1" (model card). Hugging Face. huggingface.co/...mxbai-embed-large-v1
- ^1 ^2 ^3 ^4NovaSearch. "NovaSearch/stella_en_400M_v5" (model card). Hugging Face. huggingface.co/...stella_en_400M_v5
- ^1 ^2 ^3IBM. "ibm-granite/granite-embedding-english-r2" (model card). Hugging Face. huggingface.co/...granite-embedding-english-r2
- ^1 ^2 ^3IBM. "ibm-granite/granite-embedding-311m-multilingual-r2" (model card). Hugging Face. huggingface.co/...e-embedding-311m-multilingual-r2
- ^1 ^2Voyage AI. "MongoDB/voyage-4-nano" (model card). Hugging Face. huggingface.co/...voyage-4-nano
- ^1 ^2Perplexity. "perplexity-ai/pplx-embed-v1-0.6b" (model card). Hugging Face. huggingface.co/...pplx-embed-v1-0.6b
- ^1 ^2NVIDIA. "nvidia/Nemotron-3-Embed-8B-BF16" (model card). Hugging Face. huggingface.co/...Nemotron-3-Embed-8B-BF16
- ^OpenMDW. "OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)". openmdw.ai. openmdw.ai/...1-1
- ^Google. "Gemma Terms of Use". Gemma Documentation, last modified April 1, 2026. ai.google.dev/...terms
- ^BAAI. "BAAI/bge-multilingual-gemma2" (model card). Hugging Face. huggingface.co/...bge-multilingual-gemma2
- ^1 ^2 ^3NVIDIA. "nvidia/llama-nemotron-embed-1b-v2" (model card). Hugging Face. huggingface.co/...llama-nemotron-embed-1b-v2
- ^1 ^2 ^3NVIDIA. "nvidia/llama-nemotron-rerank-1b-v2" (model card and commit history). Hugging Face. huggingface.co/...llama-nemotron-rerank-1b-v2
- ^1 ^2Tencent. "LICENSE.txt", tencent/KaLM-Embedding-Gemma3-12B-2511. Hugging Face. huggingface.co/...LICENSE.txt
- ^1 ^2 ^3ViDoRe. "vidore/colpali-v1.3" (model card). Hugging Face. huggingface.co/...colpali-v1.3
- ^Jina AI. "jinaai/jina-embeddings-v5-text-small" (model card). Hugging Face. huggingface.co/...jina-embeddings-v5-text-small
- ^1 ^2 ^3 ^4 ^5Jina AI. "jinaai/jina-embeddings-v4" (model card, LICENSE file and commit history). Hugging Face. huggingface.co/...jina-embeddings-v4
- ^1 ^2Jina AI. "jinaai/jina-reranker-v3" (model card). Hugging Face. huggingface.co/...jina-reranker-v3
- ^1 ^2Jina AI. "jinaai/jina-colbert-v2" (model card). Hugging Face. huggingface.co/...jina-colbert-v2
- ^Jina AI. "jinaai/jina-clip-v2" (model card). Hugging Face. huggingface.co/...jina-clip-v2
- ^1 ^2Jina AI. "jinaai/jina-code-embeddings-1.5b" (model card). Hugging Face. huggingface.co/...jina-code-embeddings-1.5b
- ^Salesforce. "Salesforce/SFR-Embedding-Mistral" (model card). Hugging Face. huggingface.co/...SFR-Embedding-Mistral
- ^Linq AI Research. "Linq-AI-Research/Linq-Embed-Mistral" (model card). Hugging Face. huggingface.co/...Linq-Embed-Mistral
- ^1 ^2 ^3ViDoRe. "vidore/colqwen2.5-v0.2" (model card). Hugging Face. huggingface.co/...colqwen2.5-v0.2
- ^1 ^2LightOn. "lightonai/Reason-ModernColBERT" (model card). Hugging Face. huggingface.co/...Reason-ModernColBERT
- ^1 ^2ZeroEntropy. "zeroentropy/zerank-2-reranker" (model card and commit history). Hugging Face. huggingface.co/...zerank-2-reranker
- ^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
- ^Nomic AI. "nomic-ai/nomic-embed-code" (model card). Hugging Face. huggingface.co/...nomic-embed-code
- ^Qwen. "Qwen/Qwen3-VL-Embedding-8B" (model card). Hugging Face. huggingface.co/...Qwen3-VL-Embedding-8B
- ^Tencent. "tencent/WeMM-Embedding-9B" (model card and LICENSE file). Hugging Face. huggingface.co/...WeMM-Embedding-9B
- ^Google. "google/siglip2-so400m-patch14-384" (model card). Hugging Face. huggingface.co/...siglip2-so400m-patch14-384
- ^OpenAI. "openai/clip-vit-large-patch14" (model card). Hugging Face. huggingface.co/...clip-vit-large-patch14
- ^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
- ^Nomic AI. "nomic-ai/colnomic-embed-multimodal-7b" (model card). Hugging Face. huggingface.co/...colnomic-embed-multimodal-7b
- ^ViDoRe. "vidore/colqwen2-v1.0" (model card). Hugging Face. huggingface.co/...colqwen2-v1.0
- ^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
- ^1 ^2 ^3ColBERT. "colbert-ir/colbertv2.0" (model card). Hugging Face. huggingface.co/...colbertv2.0
- ^LightOn. "lightonai/GTE-ModernColBERT-v1" (model card). Hugging Face. huggingface.co/...GTE-ModernColBERT-v1
- ^Answer.AI. "answerdotai/answerai-colbert-small-v1" (model card). Hugging Face. huggingface.co/...answerai-colbert-small-v1
- ^Mixedbread. "mixedbread-ai/mxbai-edge-colbert-v0-32m" (model card). Hugging Face. huggingface.co/...mxbai-edge-colbert-v0-32m
- ^1 ^2Sentence Transformers. "SentenceTransformers Documentation". sbert.net. sbert.net
- ^1 ^2Sentence Transformers. "cross-encoder/ms-marco-MiniLM-L6-v2" (model card). Hugging Face. huggingface.co/...ms-marco-MiniLM-L6-v2
- ^BAAI. "BAAI/bge-reranker-v2-m3" (model card). Hugging Face. huggingface.co/...bge-reranker-v2-m3
- ^Qwen. "Qwen/Qwen3-Reranker-8B" (model card). Hugging Face. huggingface.co/...Qwen3-Reranker-8B
- ^Alibaba. "Alibaba-NLP/gte-reranker-modernbert-base" (model card). Hugging Face. huggingface.co/...gte-reranker-modernbert-base
- ^Mixedbread. "mixedbread-ai/mxbai-rerank-large-v2" (model card). Hugging Face. huggingface.co/...mxbai-rerank-large-v2
- ^Jina AI. "jinaai/jina-reranker-v2-base-multilingual" (model card). Hugging Face. huggingface.co/...jina-reranker-v2-base-multilingual
- ^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
- ^Microsoft. "MS MARCO". microsoft.github.io/msmarco
- ^Hugging Face. "huggingface/sentence-transformers". GitHub repository, figures as of November 2026. github.com/...sentence-transformers
- ^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
- ^1 ^2Ollama. "Embeddings". Ollama documentation. docs.ollama.com/...embeddings
- ^Ollama. "ollama/ollama". GitHub repository, figures as of November 2026. github.com/...ollama
- ^1 ^2 ^3ggml-org. "LLaMA.cpp HTTP Server" (tools/server/README.md). GitHub, ggml-org/llama.cpp. github.com/...README.md
- ^ggml-org. "ggml-org/llama.cpp". GitHub repository, figures as of November 2026. github.com/...llama.cpp
- ^1 ^2vLLM project. "Pooling Models". vLLM documentation (latest developer preview). docs.vllm.ai/...pooling_models
- ^vLLM project. "vllm-project/vllm". GitHub repository, figures as of November 2026. github.com/...vllm
- ^michaelfeil. "michaelfeil/infinity". GitHub repository (README and LICENSE), figures as of November 2026. github.com/...infinity
- ^Qdrant. "qdrant/fastembed". GitHub repository (README and LICENSE), figures as of November 2026. github.com/...fastembed
- ^Hugging Face. "HfApi Client" (the created_at field of ModelInfo). huggingface_hub documentation. huggingface.co/...hf_api
- ^Jina AI. "Jina Embeddings v3: A Frontier Multilingual Embedding Model". Jina AI News, September 18, 2024. jina.ai/...a-frontier-multilingual-embedding-model
- ^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