🔎 KURE: Korea University REtrieval models

KURE is a family of Korean-English retrieval embedding models developed by the NLP & AI Lab and the HIAI Institute at Korea University.

Update Logs

2026.10.20: 🤗 KURE-v3 released with API.
2026.10.12: 🤗 KURE-Sparse-base and 🤗 KURE-Sparse-nano released.
2026.10.01: 🤗 KURE-Reranker-base and 🤗 KURE-Reranker-nano released.
2026.08.29: 🤗 KURE-v2 released.
2024.12.21: 🤗 KURE-v1 released with the MTEB-ko-retrieval leaderboard.
2024.10.02: 🤗 KoE5 and 🤗 ko-triplet-v1.0 released.

Models

Model	Type	Params	Base model
KURE-v3	Dense (single-vector)	1.1B	nvidia/Nemotron-3-Embed-1B-BF16
KURE-Sparse-base	Learned Sparse Retrieval (LSR)	593M	omai-research/milco-650m
KURE-Sparse-nano	Learned Sparse Retrieval (LSR)	149M	nlpai-lab/KURE-v2-unsupervised
KURE-Reranker-base	Reranker (cross-encoder)	1.7B	Qwen/Qwen3-1.7B
KURE-Reranker-nano	Reranker (cross-encoder)	149M	skt/A.X-Encoder-base
KURE-v2	Late-interaction (multi-vector)	154M	skt/A.X-Encoder-base
KURE-v1	Dense (single-vector)	568M	BAAI/bge-m3
KoE5	Dense (single-vector)	560M	intfloat/multilingual-e5-large

Environment

We use uv to manage the environment.

uv sync                                                # dense, sparse and cross-encoder models
uv sync --extra late-interaction --no-default-groups   # KURE-v2 with PyLate

Usage

API

KURE-v3, KURE-Sparse-{base, nano} and KURE-Reranker-nano are served at kure.kullm.ai. Sign in with GitHub there to get an API key. Each request takes up to 8 texts (100 documents for reranking).

Endpoint	Models
`/embeddings`	`kure-v3`
`/sparse-embeddings`	`kure-sparse-base`, `kure-sparse-nano`
`/rerank`	`kure-reranker-nano`

import os

import requests
from openai import OpenAI

BASE_URL = "https://kure.kullm.ai/api/v1"
API_KEY = os.environ["KURE_API_KEY"]

query = "훈민정음은 언제 만들어졌나요?"
documents = [
    "세종대왕은 1443년에 훈민정음을 창제하고 1446년에 이를 반포하였다.",
    "김치는 배추나 무를 소금에 절인 뒤 고춧가루와 젓갈을 넣어 발효시킨 음식이다.",
    "한라산은 해발 1,947m로 남한에서 가장 높은 산이며 제주도 중앙에 자리한다.",
]

# KURE-v3 (OpenAI-compatible). input_type adds the query or document prompt. Vectors are 2048-d, pass dimensions=(256 to 2048) for a smaller Matryoshka size
client = OpenAI(base_url=BASE_URL, api_key=API_KEY)
query_embedding = client.embeddings.create(
    model="kure-v3", input=query, extra_body={"input_type": "query"}
)
document_embeddings = client.embeddings.create(
    model="kure-v3", input=documents, extra_body={"input_type": "document"}
)

headers = {"Authorization": f"Bearer {API_KEY}"}

# KURE-Sparse returns vocabulary ids (indices) and their weights (values)
sparse = requests.post(
    f"{BASE_URL}/sparse-embeddings",
    headers=headers,
    json={"model": "kure-sparse-nano", "input": documents, "input_type": "document"},
).json()

# KURE-Reranker returns the documents sorted by relevance_score (0 to 1, higher is more relevant)
ranked = requests.post(
    f"{BASE_URL}/rerank",
    headers=headers,
    json={"model": "kure-reranker-nano", "query": query, "documents": documents},
).json()

KURE-Dense

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("nlpai-lab/KURE-v3", model_kwargs={"attn_implementation": "flash_attention_2"})
# model = SentenceTransformer("nlpai-lab/KURE-v3", model_kwargs={"attn_implementation": "flash_attention_2"}, truncate_dim=1024)  # smaller Matryoshka vectors

# KURE-v2 is late-interaction and needs sentence-transformers>=6.0:
# from sentence_transformers import MultiVectorEncoder
# model = MultiVectorEncoder("nlpai-lab/KURE-v2")

# model = SentenceTransformer("nlpai-lab/KURE-v1")
# model = SentenceTransformer("nlpai-lab/KoE5", prompts={"query": "query: ", "document": "passage: "})

query = "훈민정음은 언제 만들어졌나요?"
documents = [
    "세종대왕은 1443년에 훈민정음을 창제하고 1446년에 이를 반포하였다.",
    "김치는 배추나 무를 소금에 절인 뒤 고춧가루와 젓갈을 넣어 발효시킨 음식이다.",
    "한라산은 해발 1,947m로 남한에서 가장 높은 산이며 제주도 중앙에 자리한다.",
]

# encode_query / encode_document add each model's query and document prompts
query_embeddings = model.encode_query(query)
document_embeddings = model.encode_document(documents)

# Higher is more relevant (cosine, or MaxSim for KURE-v2)
scores = model.similarity(query_embeddings, document_embeddings)

KURE-Sparse

from sentence_transformers import SparseEncoder

model = SparseEncoder("nlpai-lab/KURE-Sparse-nano", trust_remote_code=True)
# model = SparseEncoder("nlpai-lab/KURE-Sparse-base", trust_remote_code=True)

query = "훈민정음은 언제 만들어졌나요?"
documents = [
    "세종대왕은 1443년에 훈민정음을 창제하고 1446년에 이를 반포하였다.",
    "김치는 배추나 무를 소금에 절인 뒤 고춧가루와 젓갈을 넣어 발효시킨 음식이다.",
    "한라산은 해발 1,947m로 남한에서 가장 높은 산이며 제주도 중앙에 자리한다.",
]

# encode_query / encode_document add the query and document markers KURE-Sparse-nano expects
query_embeddings = model.encode_query(query)
document_embeddings = model.encode_document(documents)

# Dot-product scores (higher is more relevant)
scores = model.similarity(query_embeddings, document_embeddings)

KURE-Reranker

from sentence_transformers import CrossEncoder

model = CrossEncoder("nlpai-lab/KURE-Reranker-nano", max_length=8192)
# model = CrossEncoder("nlpai-lab/KURE-Reranker-base", max_length=8192)

query = "훈민정음은 언제 만들어졌나요?"
documents = [
    "세종대왕은 1443년에 훈민정음을 창제하고 1446년에 이를 반포하였다.",
    "김치는 배추나 무를 소금에 절인 뒤 고춧가루와 젓갈을 넣어 발효시킨 음식이다.",
    "한라산은 해발 1,947m로 남한에서 가장 높은 산이며 제주도 중앙에 자리한다.",
]

# Higher is more relevant
results = model.rank(query, documents, return_documents=True)

Evaluation

We evaluate on the nine MTEB(kor, v2) Retrieval tasks with the latest mteb (v2), using the Korean subsets only (Belebele: kor_Hang-kor_Hang; MLDR is the mean of its dev/test splits):

Ko-StrategyQA: Korean ODQA multi-hop retrieval (translated StrategyQA)
AutoRAGRetrieval: document retrieval over parsed PDFs in finance, public sector, healthcare, law, and commerce
MIRACLRetrieval: Wikipedia-based retrieval, ~1.5M documents
PublicHealthQA: medical / public-health domain retrieval
BelebeleRetrieval: FLORES-200-based retrieval
MrTidyRetrieval: Wikipedia-based retrieval, ~1.5M documents
MultiLongDocRetrieval: long-document retrieval across domains
LawIRKo: Korean statute / case-law retrieval
SQuADKorV1Retrieval: KorQuAD v1 passage retrieval

Dense Leaderboard

Dense models are evaluated with mteb, and late-interaction models with mteb + PLAID (4-bit) retrieval. Models with results in the official MTEB results repository use those scores, and the others are run with eval/dense/evaluate.py. To evaluate a model, pass its name and paradigm:

uv run python eval/dense/evaluate.py nlpai-lab/KURE-v3 single-vector
uv run --extra late-interaction --no-default-groups python eval/dense/evaluate.py nlpai-lab/KURE-v2 multi-vector

Average over the nine tasks. Per-task scores are in eval/dense/results/README.md and on the official MTEB Leaderboard.

Model	Type	Params	Avg nDCG@10	Avg Recall@10
nlpai-lab/KURE-v2	Late-interaction	154M	0.8160	0.8921
nvidia/Nemotron-3-Embed-8B-BF16	Dense	8.0B	0.8127	0.8934
yjoonjang/colbert-ko-en-v2	Late-interaction	149M	0.8063	0.8827
nlpai-lab/KURE-v3	Dense	1.1B	0.8027	0.8857
sionic-ai/comsat-embed-ko-8b-preview	Dense	7.6B	0.7927	0.8876
lightonai/mLateOn	Late-interaction	307M	0.7906	0.8740
Qwen/Qwen3-Embedding-8B	Dense	7.6B	0.7826	0.8815
nvidia/Nemotron-3-Embed-1B-BF16	Dense	1.1B	0.7817	0.8690
Qwen/Qwen3-Embedding-4B	Dense	4.0B	0.7737	0.8753
perplexity-ai/pplx-embed-v1-4b	Dense	4.0B	0.7705	0.8699
microsoft/harrier-oss-v1-27b	Dense	27.0B	0.7667	0.8624
dragonkue/snowflake-arctic-embed-l-v2.0-ko	Dense	568M	0.7653	0.8609
codefuse-ai/F2LLM-v2-8B	Dense	7.6B	0.7638	0.8593
telepix/PIXIE-Rune-v1.5	Dense	568M	0.7618	0.8596
nlpai-lab/KURE-v1	Dense	568M	0.7616	0.8629
kakaocorp/kanana-2-3b-embedding	Dense	3.2B	0.7584	0.8569
nvidia/llama-nemotron-embed-1b-v2	Dense	1.2B	0.7564	0.8560
dragonkue/BGE-m3-ko	Dense	568M	0.7547	0.8513
BAAI/bge-m3	Dense	568M	0.7509	0.8588
jinaai/jina-embeddings-v5-text-small-retrieval	Dense	596M	0.7419	0.8434
perplexity-ai/pplx-embed-v1-0.6b	Dense	596M	0.7390	0.8377
perplexity-ai/pplx-embed-v1-late-0.6b	Late-interaction	596M	0.7381	0.8328
nlpai-lab/KoE5	Dense	560M	0.7337	0.8300
nlpai-lab/KURE-v2-unsupervised	Late-interaction	154M	0.7283	0.8268
google/embeddinggemma-300m	Dense	308M	0.7254	0.8383
kakaocorp/kanana-2-1.3b-embedding	Dense	1.3B	0.7172	0.8204
Qwen/Qwen3-Embedding-0.6B	Dense	596M	0.7140	0.8222
dragonkue/colbert-ko-0.1b	Late-interaction	149M	0.6776	0.7723
yjoonjang/colbert-ko-v1	Late-interaction	149M	0.6282	0.7212

Truncated to 1,024 dimensions, KURE-v3 keeps 99% of its full-size score and still outperforms comsat-embed-ko-8b-preview and Qwen3-Embedding-8B (4,096 dimensions). Per-dimension scores are in eval/dense/results/README.md.

RTEB (eng)

Dense and late-interaction models are also evaluated on RTEB, an English retrieval benchmark, using its eight public tasks outside the code domain. The average differs from the RTEB leaderboard score, which also covers private and code tasks. Per-task scores are in eval/dense/results/README.md.

uv run python eval/dense/evaluate.py nlpai-lab/KURE-v3 single-vector --suite rteb_eng

Sparse Leaderboard

Learned sparse models are evaluated on the same nine tasks under the same conventions, with mteb's SparseEncoder wrapper in bf16 at up to 8,192 tokens and without pruning the vectors. Next to nDCG@10, we report the average number of active dimensions per query and per document, and FLOPs. To evaluate a sparse model:

uv run python eval/sparse/evaluate.py nlpai-lab/KURE-Sparse-nano

Results are averaged over the nine Korean retrieval tasks. The dashed lines mark dense models for reference. Per-task scores, along with English BEIR and RTEB results, are in eval/sparse/results/README.md.

Model	Params	Avg nDCG@10	Query dims	Document dims	FLOPs
nlpai-lab/KURE-Sparse-base	593M	0.7984	100	1,897	36.5
omai-research/milco-650m	593M	0.7824	105	1,167	32.3
nlpai-lab/KURE-Sparse-nano	149M	0.7716	147	843	23.0
telepix/PIXIE-Splade-v1.5	149M	0.7608	117	1,541	37.3
yjoonjang/splade-ko-v1	149M	0.7526	109	742	29.1
telepix/PIXIE-Splade-v1.0	149M	0.6636	121	625	22.4
opensearch-project/opensearch-neural-sparse-encoding-multilingual-v1	167M	0.5444	29	176	6.8

English Retrieval (BEIR, RTEB)

Sparse models are also evaluated on two English benchmarks with nDCG@10: the 13 BEIR tasks (BEIR-13), and the eight public English tasks of RTEB outside the code domain. SparseUp runs at its own length limits (128 query and 512 document tokens), and the other models at up to 8,192 tokens. The RTEB average differs from the RTEB leaderboard score, which also covers private and code tasks.

uv run python eval/sparse/evaluate.py nlpai-lab/KURE-Sparse-base --suite beir_eng --max-seq-length 8192
uv run python eval/sparse/evaluate.py nlpai-lab/KURE-Sparse-base --suite rteb_eng --max-seq-length 8192

Per-task scores are in eval/sparse/results/README.md.

Reranker Leaderboard

Rerankers are evaluated on the same nine tasks with the protocol of reranker-simple-benchmark: each query's candidates are its gold documents plus, as negatives, the BM25 top-50 minus the golds, every model runs in bf16 at up to 8,192 tokens, and we report nDCG@10 with throughput in pairs per second (PPS), measured on one NVIDIA RTX A6000 48GB. To evaluate a reranker:

bash eval/cross-encoder/fetch_eval_assets.sh   # once: the BM25 candidate pools
uv run python eval/cross-encoder/evaluate.py nlpai-lab/KURE-Reranker-nano --speed

Results are averaged over the nine Korean retrieval tasks. Per-task nDCG@10 and PPS are in eval/cross-encoder/results/README.md.

Training Details

KURE-v3

Training code in train/dense.

Fine-tuned from nvidia/Nemotron-3-Embed-1B-BF16 by distilling a teacher's scores over each query's candidates on the KURE-v2 SFT data (min-max normalized KL), up to 8,192 tokens, 1 epoch.
Matryoshka training at 2,048, 1,024, 512 and 256 dimensions.

KURE-Sparse

Training code in train/sparse.

Distilled from Qwen3-Reranker teacher on the KURE-v2 SFT data with an in-batch contrastive term plus a KL term over each query's candidates, an L1 sparsity regularization on queries and documents, up to 8,192 tokens, 1 epoch.
KURE-Sparse-base: fine-tuned from omai-research/milco-650m, scoring over 30,522 English pivot terms plus 250,002 source-language tokens.
KURE-Sparse-nano: the A.X-Encoder-base MLM head grafted onto KURE-v2-unsupervised, pre-finetuned on 20.7M pairs, then fine-tuned. Resulting in 36,282 vocabulary terms.

KURE-Reranker

Training code in train/cross-encoder.

Distilled from a Qwen3-Reranker teacher with pointwise MSE on the KURE-v2 SFT data flattened to 33.3M query-document pairs, 1 epoch.
KURE-Reranker-base: Qwen/Qwen3-1.7B scored as logit("yes") − logit("no") over the Qwen3-Reranker template, up to 8,192 tokens.
KURE-Reranker-nano: skt/A.X-Encoder-base with a single-logit head, up to 8,192 tokens.

KURE-v2

The full two-stage training code is in train/late-interaction.

Built on skt/A.X-Encoder-base with a multi-layer projection head (128-d per token).
Stage 1 (PFT): large in-batch contrastive learning on 20.7M weakly related Korean/English pairs, released as KURE-v2-unsupervised.
Stage 2 (SFT): contrastive learning + KL distillation from a reranker teacher on 3.03M triplets with hard negatives and false-negative filtering.

KURE-v1

Fine-tuned from BAAI/bge-m3 on ~2M Korean query-document-hard-negative(5) pairs.
CachedGISTEmbedLoss, batch size 4,096, lr 2e-5, 1 epoch.

KoE5

Fine-tuned from intfloat/multilingual-e5-large on ko-triplet-v1.0 (~700K+ examples).
CachedMultipleNegativesRankingLoss, batch size 512, lr 1e-5, 1 epoch. Requires query: / passage: prefixes.

Serving

Index-size and latency benchmarks of KURE-v2 across ANN backends, token pooling and binary quantization, with runnable configurations, are in deploy/late-interaction.

License

MIT

Citation

If you find our models helpful, please consider citing:

@misc{kure-v3,
  title  = {KURE-v3: a Korean-English bilingual dense retriever},
  author = {Jang, Youngjoon and Son, Junyoung and Lee, Taemin and Hong, Seongtae and Lee, Jungseob and Lim, Heuiseok},
  year   = {2026},
  url    = {https://huggingface.co/nlpai-lab/KURE-v3},
}

@misc{kure-reranker,
  title  = {KURE-Reranker: Korean-English bilingual reranking models},
  author = {Jang, Youngjoon and Hong, Seongtae and Son, Junyoung and Lee, Taemin and Lim, Heuiseok},
  year   = {2026},
  url    = {https://huggingface.co/nlpai-lab/KURE-Reranker-base},
}

@misc{kure-v2,
  title  = {KURE-v2: a Korean-English bilingual late-interaction retriever},
  author = {Jang, Youngjoon and Son, Junyoung and Lee, Taemin and Hong, Seongtae and Lim, Heuiseok},
  year   = {2026},
  url    = {https://huggingface.co/nlpai-lab/KURE-v2},
}

@inproceedings{jang2025kure,
  title={KURE: Embedding Model for Korean-Specific Retrieval},
  author={Jang, Youngjoon and Son, Junyoung and Lee, Taemin and Hong, Seongtae and Park, JeongBae and Lim, Heuiseok},
  booktitle={Annual Conference on Human and Language Technology},
  pages={129--134},
  year={2025},
  organization={Human and Language Technology}
}

@inproceedings{jang2024koe5,
  title={KoE5: A New Dataset and Model for Improving Korean Embedding Performance},
  author={Jang, Youngjoon and Son, Junyoung and Park, Chanjun and Choi, Soonwoo and Lee, Byeonggoo and Lee, Taemin and Lim, Heuiseok},
  booktitle={Annual Conference on Human and Language Technology},
  pages={239--244},
  year={2024},
  organization={Human and Language Technology}
}

Name		Name	Last commit message	Last commit date
assets		assets
deploy/late-interaction		deploy/late-interaction
eval		eval
train		train
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
README_ko.md		README_ko.md
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Navigation Menu

Repository navigation

About

Repository files navigation

🔎 KURE: Korea University REtrieval models

Update Logs

Models

Environment

Usage

API

KURE-Dense

KURE-Sparse

KURE-Reranker

Evaluation

Dense Leaderboard

RTEB (eng)

Sparse Leaderboard

English Retrieval (BEIR, RTEB)

Reranker Leaderboard

Training Details

KURE-v3

KURE-Sparse

KURE-Reranker

KURE-v2

KURE-v1

KoE5

Serving

License

Citation

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

🔎 KURE: Korea University REtrieval models

Update Logs

Models

Environment

Usage

API

KURE-Dense

KURE-Sparse

KURE-Reranker

Evaluation

Dense Leaderboard

RTEB (eng)

Sparse Leaderboard

English Retrieval (BEIR, RTEB)

Reranker Leaderboard

Training Details

KURE-v3

KURE-Sparse

KURE-Reranker

KURE-v2

KURE-v1

KoE5

Serving

License

Citation

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages