🔎 KURE: Korea University Retrieval Embedding models

English | 한국어

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

• 2026.08.29: 🤗 KURE-v2 released: a Korean-English bilingual late-interaction (multi-vector) model, state of the art on the MTEB(kor, v2) retrieval benchmark.
• 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-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

KURE-v2 encodes every token into a 128-dimensional vector and scores query-document pairs with MaxSim, which preserves token-level semantics that single-vector models compress away. It supports documents up to 8,192 tokens and needs no instruction prefixes.

Environment

We use uv to manage the environment.

uv sync

Usage

KURE-v2 (late-interaction)

PyLate

uv add pylate

from pylate import indexes, models, retrieve

model = models.ColBERT(model_name_or_path="nlpai-lab/KURE-v2")

index = indexes.PLAID(index_folder="pylate-index", index_name="index", override=True)

documents_ids = ["1", "2", "3"]
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.",
]

documents_embeddings = model.encode(documents, is_query=False)
index.add_documents(documents_ids=documents_ids, documents_embeddings=documents_embeddings)

retriever = retrieve.ColBERT(index=index)
queries_embeddings = model.encode(["훈민정음은 언제 만들어졌나요?"], is_query=True)
scores = retriever.retrieve(queries_embeddings=queries_embeddings, k=10)

Sentence-Transformers

uv add sentence-transformers

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("nlpai-lab/KURE-v2")

query = "훈민정음은 언제 만들어졌나요?"
documents = [
    "세종대왕은 1443년에 훈민정음을 창제하고 1446년에 이를 반포하였다.",
    "김치는 배추나 무를 소금에 절인 뒤 고춧가루와 젓갈을 넣어 발효시킨 음식이다.",
]

query_embeddings = model.encode_query(query)
document_embeddings = model.encode_document(documents)

# MaxSim late-interaction scoring (higher is more relevant)
scores = model.similarity(query_embeddings, document_embeddings)

KURE-v1 / KoE5 (dense)

Sentence-Transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("nlpai-lab/KURE-v1")
# model = SentenceTransformer("nlpai-lab/KoE5")  # needs "query: " / "passage: " prefixes

embeddings = model.encode(["첫 번째 문장", "두 번째 문장"])
similarities = model.similarity(embeddings, embeddings)

Evaluation

We evaluate on the nine MTEB(kor, v2) Retrieval tasks with the latest mteb (v2.x), 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

Late-interaction models are measured directly with mteb + PLAID retrieval; dense rows come from the official MTEB results repository. To evaluate a model yourself, pass its name and paradigm:

uv run python eval/evaluate.py nlpai-lab/KURE-v2 multi-vector
uv run python eval/evaluate.py nlpai-lab/KURE-v1 single-vector

The per-task result files backing the leaderboard are in eval/results, and the table below is regenerated from them:

uv run python eval/make_leaderboard.py

MTEB-ko-retrieval Leaderboard

Average over the nine tasks. Full per-task results are 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
yjoonjang/colbert-ko-en-v2	Late-interaction	149M	0.8063	0.8827
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
Qwen/Qwen3-Embedding-4B	Dense	4.0B	0.7737	0.8753
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
dragonkue/BGE-m3-ko	Dense	568M	0.7547	0.8513
BAAI/bge-m3	Dense	568M	0.7509	0.8588
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
dragonkue/colbert-ko-0.1b	Late-interaction	149M	0.6776	0.7723
yjoonjang/colbert-ko-v1	Late-interaction	149M	0.6282	0.7212

Training Details

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, MaxSim scoring); trained with PyLate.
• 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

Training code in train/dense.

• 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

KURE-v2 is a late-interaction model: each document is stored as a set of token vectors, so the practical questions for deployment are index size and search cost. We benchmarked KURE-v2 across ANN backends and compression schemes on the 9 Korean MTEB retrieval tasks, against five single-vector baselines served with faiss HNSW. All numbers are end-to-end: batch-1 query encoding + index search, measured serially on one A100 80GB.

Runnable versions of these configurations are in deploy/late-interaction (uv run deploy/late-interaction/run.py --index plaid).

Two things the figures show:

• Hierarchical token pooling (x2) halves the index (24.2 -> 12.5 GB) with no measurable nDCG loss. Asymmetric binary quantization (1-bit document tokens, bf16 queries) shrinks it 4.8x for 0.98. Stacking the two (pooling x3 + binary), the entire 9-corpus index fits in 1.7 GB, smaller than every single-vector HNSW index (7.0-25.4 GB, fp16 vectors), while still outscoring the best single-vector model.
• A live query arrives as text: 4B-8B single-vector models spend 38-40 ms encoding it, capping them at ~25 QPS no matter how fast HNSW is. KURE-v2 encodes in 13.8 ms (154M params), so every configuration except MUVERA serves 44-57 QPS, roughly 2x the 8B single-vector models, at higher quality.

Large corpora: tail latency

On the largest corpus (MIRACL, ~1.5M documents) an exhaustive 1-bit scan costs O(corpus): p95 climbs to 156 ms, and pooling the tokens 3x only brings it to 74 ms. Generating candidates with faiss BinaryIVF (Hamming search over the same 1-bit index) and re-scoring them with exact asymmetric MaxSim cuts p95 to 38 ms on the same 2.2 GB index, lower tail latency than the 4B-8B single-vector baselines (42 ms) at higher nDCG. For large collections, use a candidate-generating index (PLAID or BinaryIVF), not an exhaustive scan.

Measurement details

• Hardware: 1x NVIDIA A100 80GB, 2x AMD EPYC 7513 (64 cores), 1.2 TB RAM.
• Software: faiss-cpu 1.15.0, fast-plaid 1.6.0, sentence-transformers 6.0.0, PyTorch 2.8.0.
• Protocol: batch-1, serial. Index-search latency: 10 warmup queries, then every query of the task measured once (QPS = 1/mean). Query-encoding latency: 5 warmup, 50 measured. End-to-end = encoding + search.
• Precision: encoding in bf16; each index stores its own format (HNSW fp16 vectors, PLAID 4-bit residuals, binary 1-bit).
• Index size: the full serialized index on disk (vectors, graph, codebooks; external doc-id mapping excluded). PLAID indexes are frozen (fast-plaid freeze()): the merged search-time codes/residuals only, without the per-shard build copies or the raw embeddings fast-plaid keeps for corpora of <= 1,000 documents.
• Tasks: the 9 Korean MTEB retrieval tasks; MLDR is the mean of its dev/test splits; nDCG@10 x100.
• HNSW: IndexHNSWSQ with fp16-stored vectors (inner product on L2-normalized embeddings; lossless for the bf16 embeddings), M=32, efConstruction=200, efSearch=64.
• PLAID: nbits=4, all other settings fast-plaid defaults (kmeans_niters=4, n_ivf_probe=8, n_full_scores=4096). nbits=2/1 give 14.2/9.1 GB at 81.40/80.70 nDCG.
• MUVERA: num_repetitions=10, num_simhash_projections=6, final_projection_dimension=8192, exact-MaxSim rerank of the top 1,000.
• BinaryIVF: nlist=floor(sqrt(total tokens)) capped at 65,536, nprobe=32, top-128 Hamming tokens per query token, exact asymmetric-MaxSim rerank of the top 1,000 documents.
• Token pooling: hierarchical (Ward linkage), pool_factor 2-3, documents only.

License

MIT

Citation

If you find our models helpful, please consider citing:

@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}
}