KURE-v2 adds multi-vector retrieval at 154M params: 128d per token with MaxSim late interaction

tomaarsen · x · 2026-10-23

tomaarsen (sentence-transformers maintainer) walks through nlpai-lab's Korean KURE retrieval family, where KURE-v2 completes the set with MultiVectorEncoder support.

• Just 154M parameters: keeps a 128d vector per token and uses MaxSim, letting different query tokens match different parts of a document (late interaction / ColBERT-style)
• The thread also covers KURE-Sparse-base/nano (593M/149M), which learn weighted terms searchable via an inverted index, reporting 79.84 / 77.16 nDCG@10 across nine Korean retrieval tasks
• Both load with SparseEncoder and trustremotecode=True; models are on Hugging Face (Apache-2.0)

Related event: Korea University open-sources KURE, a full retrieval family spanning all four Sentence Transformers model types(8 posts)→

Original post →

More from Models

• Why frontier AI labs' bug bounties are cheap: their models flood the supply — lxrjl · 2026-10-23
• 7 OpenAI updates this week that cut your API bill, from 12-hour caching to Decisions API — MrCashMahon · 2026-10-23
• 100 classic human-written texts stress-test Pangram AI detector: formatting alone triggers false positives — GolinoHudson · 2026-10-23
• IQ2_XXS quant + 64k context makes a 27B model spin forever instead of coding — mailto_devnull · 2026-10-23
• UK AISI: Mythos 5 ignored DO NOT ATTACK warnings, delivered malware in 60% of tests — Hesamation · 2026-10-23
• Decision models are shockingly sensitive to question order and random noise — antoine_chaffin · 2026-10-23

Models channel →