🔎 KURE-v2

KURE-v2 is a Korean-English bilingual late-interaction (multi-vector) retrieval model built on skt/A.X-Encoder-base. It encodes every token into a 128-dimensional vector and scores query–document pairs with MaxSim.

At 154M parameters it averages nDCG@10 0.8160 across the nine MTEB(kor, v2) retrieval tasks — the strongest model on that suite, above every single-vector model measured, including one 175× its size.

It is trained in two stages: unsupervised contrastive pretraining on 20.7M unlabeled pairs (KURE-v2-unsupervised), then supervised fine-tuning with contrastive loss and KL distillation from a reranker.

Key Characteristics

• Late interaction: one 128-d vector per token, scored with MaxSim.
• Long documents: up to 8,192 tokens.
• Compact: 154M parameters.
• No instruction prefixes: queries and documents need no task instruction. Query expansion to 64 tokens is handled by the model.

Usage

pip install -U 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,
    batch_size=32,
    is_query=False,
    show_progress_bar=True,
)

index.add_documents(
    documents_ids=documents_ids,
    documents_embeddings=documents_embeddings,
)

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

retriever = retrieve.ColBERT(index=index)

queries_embeddings = model.encode(
    ["훈민정음은 언제 만들어졌나요?"],
    batch_size=32,
    is_query=True,
    show_progress_bar=True,
)

scores = retriever.retrieve(
    queries_embeddings=queries_embeddings,
    k=10,
)

from pylate import models, rank

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

queries = [
    "전기차 폐배터리는 어떻게 재활용하나요?",
    "겨울에 한라산을 오를 때 필요한 장비는?",
]
documents = [
    [
        "폐배터리에서 리튬과 코발트를 회수하는 습식 제련 공정이 상용화되고 있다.",
        "급속 충전기는 30분 내외로 배터리를 80%까지 충전할 수 있다.",
    ],
    [
        "겨울 한라산 산행에는 아이젠과 방한 장갑이 필수이며 입산 시간이 제한된다.",
        "제주 올레길은 해안을 따라 이어지는 27개 코스로 구성되어 있다.",
        "적설기에는 등산화에 스패츠를 착용해 눈이 들어가는 것을 막는 것이 좋다.",
    ],
]
documents_ids = [[1, 2], [1, 3, 2]]

queries_embeddings = model.encode(queries, is_query=True)
documents_embeddings = model.encode(documents, is_query=False)

reranked_documents = rank.rerank(
    documents_ids=documents_ids,
    queries_embeddings=queries_embeddings,
    documents_embeddings=documents_embeddings,
)

pip install -U 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.",
    "한라산은 해발 1,947m로 남한에서 가장 높은 산이며 제주도 중앙에 자리한다.",
]

query_embeddings = model.encode_query(query)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings[0].shape)
# (64, 128) (29, 128)

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

Evaluation

KURE-v2 is evaluated on the nine MTEB(kor, v2) Retrieval tasks. We report nDCG@10. The results are also shown in the official MTEB Leaderboard

Late-interaction rows were measured with mteb 2.18.16 and PLAID retrieval. Single-vector rows are taken from the official MTEB results repository, except for Belebele, where only the Korean-query / Korean-corpus subset is used. The original version also includes cross-lingual subsets (Korean query – English corpus, English query – Korean corpus).

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.

Two things the figures show:

• Hierarchical token pooling (x2) halves the index for a 0.04 nDCG drop. Asymmetric binary quantization (1-bit document tokens, bf16 queries) shrinks it 9.4x for 1.05. Stacking the two (pooling x3 + binary), the entire 9-corpus index fits in 1.7 GB, smaller than every single-vector HNSW index (13.1-50.0 GB), while still outscoring the best single-vector model (79.57 vs 79.07).
• 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 43-55 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 (43 ms) at higher nDCG. For large collections, use a candidate-generating index (PLAID or BinaryIVF), not an exhaustive scan.

Citation

@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{santhanam-etal-2022-colbertv2,
  title     = {ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction},
  author    = {Santhanam, Keshav and Khattab, Omar and Saad-Falcon, Jon and Potts, Christopher and Zaharia, Matei},
  booktitle = {Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year      = {2022},
  pages     = {3715--3734},
}

@misc{PyLate,
  title  = {PyLate: Flexible Training and Retrieval for Late Interaction Models},
  author = {Chaffin, Antoine and Sourty, Raphaël},
  year   = {2024},
  url    = {https://github.com/lightonai/pylate},
}