Korea University open-sources KURE, a full retrieval family spanning all four Sentence Transformers model types

The NLP & AI Lab at Korea University (nlpai-lab) has released KURE (Korea University REtrieval models), a family of open-source retrieval models covering Korean-English bilingual use and the full pipeline from document indexing to candidate reranking. According to sentence-transformers maintainer tomaarsen, this may be the first suite to support all 4 model types in Sentence Transformers: Dense, Sparse, Multi-vector embeddings, and Rerankers—all usable through Sentence Transformers.

Confirmed

KURE-v3: a 1.1B-parameter dense embedding model producing a single 2048-dimensional vector per text, with Matryoshka truncation down to 256 dimensions; on 8 public English RTEB tasks (excluding code), it approaches Qwen3-Embedding-8B—roughly 8x its size—while retaining 99% accuracy at 1024 dimensions.
KURE-v2: 154M parameters, adds MultiVectorEncoder support, outputting a 128-dimensional vector per token scored via MaxSim, completing the family's multi-vector retrieval piece.
KURE-Sparse-base/nano: 593M / 149M variants that learn weighted terms and can be retrieved directly with an inverted index; they achieve 79.84 / 77.16 nDCG@10 across 9 Korean retrieval tasks.
Reranker-base/nano: recently released and just retrained.

Why it matters

KURE's selling point is "small but strong": the 1.1B model rivals 8B competitors on English benchmarks, and with Matryoshka truncation it offers a flexible accuracy-versus-cost trade-off.
With all four architectures on board, users can compose the full indexing-to-reranking pipeline within one ecosystem—a complete out-of-the-box solution for Korean-English bilingual retrieval.

2026-10-23 ~ 2026-10-23 · 8 related posts

Primary sources

KURE-v3 debuts: first model suite covering all 4 Sentence Transformers types — tomaarsen · 2026-10-23 20:24 (UTC+08:00)
KURE-v3: 1.1B embedding model nearly matches Qwen3-Embedding-8B, keeps 99% score at 1024d — tomaarsen · 2026-10-23 20:24 (UTC+08:00)
Korea University open-sources KURE, a full Korean-English retrieval toolkit — tomaarsen · 2026-10-23 20:25 (UTC+08:00)

[source] KURE-v3 debuts: first model suite covering all 4 Sentence Transformers types — tomaarsen · 2026-10-23
KURE suite may be first open model family covering all 4 Sentence Transformers model types — tomaarsen · 2026-10-23
KURE family roundup: 1.1B model nearly matches 8B rivals on English RTEB, tops Korean charts — tomaarsen · 2026-10-23
[source] KURE-v3: 1.1B embedding model nearly matches Qwen3-Embedding-8B, keeps 99% score at 1024d — tomaarsen · 2026-10-23
KURE-Sparse-base/nano land at 79.84/77.16 nDCG@10 on Korean retrieval, searchable via inverted index — tomaarsen · 2026-10-23
KURE-v2 adds multi-vector retrieval at 154M params: 128d per token with MaxSim late interaction — tomaarsen · 2026-10-23
KURE rerankers demoed: 149M nano model boosts the right answer to top spot — tomaarsen · 2026-10-23
[source] Korea University open-sources KURE, a full Korean-English retrieval toolkit — tomaarsen · 2026-10-23