KURE-v3
Dense1.1BnDCG@10 0.8027
A single-vector retrieval model you can use with any vector DB. It supports long documents (8K tokens) and Korean–English bilingual retrieval, and you can truncate dimensions to cut storage costs.
NEW
KURE-v3
KURE-v3 is a 1.1B dense embedding model for Korean and English retrieval. Try it with the OpenAI SDK!
- Korean retrieval
- 0.8027
- Parameters
- 1.1B
- Max tokens
- 8,192
- MRL dimensions
- 256–2048
Query고려대학교는 어디에 있나요?
- 1Answer
고려대학교는 서울특별시 성북구 안암로 145에 캠퍼스를 두고 있다.
0.596 - 2Partial answer
고려대학교 세종캠퍼스는 세종특별자치시 조치원읍에 있다.
0.436 - 3Similar topic
연세대학교는 서울특별시 서대문구 신촌동에 위치한 사립 대학이다.
0.360 - 4Keyword trap
고려는 918년 왕건이 세운 왕조로, 수도는 개경이었다.
0.233 - 5Unrelated
대학 입시에서 수시 모집 원서 접수는 보통 9월에 진행된다.
0.051
Dims
Custom →Quickstart
OpenAI-compatible API. Use input_type="query" for queries and "document" for documents.
from openai import OpenAI
client = OpenAI(base_url="https://kure.kullm.ai/api/v1", api_key="kure-...")
query = client.embeddings.create(
model="kure-v3",
input="고려대학교는 어디에 있나요?",
extra_body={"input_type": "query"},
)
docs = client.embeddings.create(
model="kure-v3",
input=["고려대학교는 서울특별시 성북구 안암로 145에 있다."],
dimensions=1024, # 256 | 512 | 768 | 1024 | 2048
extra_body={"input_type": "document"},
)KURE Family
Pick the model that fits. Models with the API badge are ready to call; every model is on Hugging Face.
KURE-v2
Late-Interaction154MnDCG@10 0.8160
When retrieval quality comes first. Multi-vector (ColBERT-style), so the index is larger.
KURE-Sparse
Sparse593M · 149MnDCG@10 0.7984
Learned sparse vectors for inverted indexes (a BM25 replacement). Great for hybrid search.
kure-sparse-nanokure-sparse-base
KURE-Reranker
Reranker1.7B · 149M
Re-rank the top results of a first-stage retriever for higher precision.
kure-reranker-nano
Earlier models
- KURE-v1Dense568MnDCG@10 0.7616BGE-M3-based dense model (Dec 2024)
- KoE5Dense560MnDCG@10 0.7337multilingual-e5-large based (Oct 2024)
Benchmark
Compare by model type. Retrievers (Dense, Late-Interaction, Sparse): average nDCG@10 over the nine MTEB(kor, v2) Retrieval tasks. Reranker: average nDCG@10 over nine Korean reranking benchmarks.
KURE-v3 ranks second among dense models, outperforming even comsat-embed-ko-8b-preview and Qwen3-Embedding-8B, seven times its size.
Need more?
Tell us if the free limits are not enough or you want to use KURE across your organization.
Higher limits
We raise per-account limits for research and projects, based on your recent usage.
Request higher limitsOrganizations & enterprise
Commercial use, organization-wide use, dedicated capacity, or on-premises deployment.
Contact usUpdates
KURE-v3 / API released
KURE-Sparse-base / nano released
KURE-Reranker-base / nano released
KURE-v2 released
KURE-v1 released with the MTEB-ko-retrieval leaderboard
KoE5 and ko-triplet-v1.0 released