Linkup releases SPARSEUP, a 149M sparse retriever hitting 56.4 nDCG@10 on BEIR-13 under Apache 2.0

Hugging Face / Linkup-Platform, corroborated by MarkTechPost and a sentence-transformers docs PR·medium signalXBlueskyLinkedInCopy link

Linkup Research published SPARSEUP on Hugging Face on September 19, a learned sparse encoder on a 149M ModernBERT backbone, Apache 2.0, which it calls the strongest public vocabulary-based sparse encoder under 150M on BEIR-13. It uses a logit shift, per-token top-12 vocabulary expansion, and case folding of byte-level BPE variants, and reaches over 97% recall in roughly 380 microseconds per query with a Seismic index. The useful part for RAG builders is the controlled comparison with backbone and data held fixed: LateOn 58.9, DenseOn 57.9, SPARSEUP 56.4, with SPARSEUP winning ArguAna and Touché and losing worst on FiQA.