🔎 KURE-v3

KURE-v3 is a Korean-English bilingual dense retrieval model. With 1.1B parameters, it averages 0.8027 nDCG@10 on MTEB(kor, v2) Retrieval. It outperforms 8B dense models such as comsat-embed-ko-8b-preview and Qwen3-Embedding-8B. It is also served through the KURE API.

Key Characteristics

• Dense (single-vector): one 2,048-dimensional vector per text, scored by cosine similarity.
• Matryoshka embeddings: trained at 2,048, 1,024, 512 and 256 dimensions, so vectors can be truncated.
• Long documents: up to 8,192 tokens.

Usage

Sentence Transformers

pip install -U sentence-transformers

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

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: " / "passage: " prompts KURE-v3 expects
query_embeddings = model.encode_query(query)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# (2048,) (3, 2048)

# Cosine similarity (higher is more relevant)
scores = model.similarity(query_embeddings, document_embeddings)
print(scores)
# tensor([[ 0.5386,  0.0152, -0.0211]])

API

KURE-v3 is served at kure.kullm.ai through an OpenAI-compatible endpoint. Sign in with GitHub there to get an API key.

import os

import numpy as np
from openai import OpenAI

client = 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.",
]

# input_type adds the query or document prompt. Vectors are 2048-d, pass dimensions=(256 to 2048) for a smaller Matryoshka size
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"})

# Vectors are L2-normalized, so the dot product is the cosine similarity
q = np.array(query_embedding.data[0].embedding)
d = np.array([e.embedding for e in document_embeddings.data])
print((d @ q).round(4))
# [ 0.5384  0.0145 -0.0207]

Evaluation

KURE-v3 is evaluated on the nine MTEB(kor, v2) Retrieval tasks. We report nDCG@10.

Matryoshka Dimensions

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).

RTEB (Retrieval Embedding Benchmark)

We also evaluate dense (single-vector) models on RTEB, an English retrieval benchmark that combines open and private datasets to measure how well models generalize to real-world data. We evaluate its eight public English tasks outside the code domain, so the average differs from the RTEB leaderboard score. We report nDCG@10.

License

KURE-v3 is released under the Apache License 2.0.

Citation

@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 Lim, Heuiseok},
  year   = {2026},
  url    = {https://huggingface.co/nlpai-lab/KURE-v3},
}

@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}
}