Linkup Open-Sources SparseUp: 149M-Parameter Sparse Retriever Tops BEIR-13

On September 17, Linkup Research (under LightOn) open-sourced SparseUp (linkup-sparseup-embed-v1, Apache 2.0), positioning it as the missing sparse piece alongside DenseOn and LateOn. The model has only 149M parameters, is built on a ModernBERT backbone trained on the same data, and scores 56.4 nDCG@10 on BEIR-13 — claimed by the team as the strongest public sparse retriever at this scale. Sentence Transformers maintainer tomaarsen endorsed it and shipped ecosystem support, sparking dense discussion on release day.

Confirmed

• Released by Linkup (research team including thibaultformal), open weights, Apache 2.0 (m1, m6, m8)
• 149M parameters, ModernBERT backbone, same backbone and training data (m1, m11, m12)
• 56.4 nDCG@10 on BEIR-13, strongest known at this size, substantially outperforming existing models (m1, m9, m12)
• Sentence Transformers added a SparseEncoder interface fully compatible with SparseUp, loadable from the Hugging Face Hub in a few lines (m7, m11)

Controversy and ecosystem context

• antoinechaffin argued that usable open-source sparse embedding models remain scarce and more would benefit the ecosystem (m2); prithivida countered that he released the SpladePP family (Apache-2.0) over two years ago with 4.5M downloads and integrations in chromaDB, haystack and other vector stores (m10)
• tomaarsen pointed to SBERT's documentation listing sparse encoder models beyond naver/splade, several under MIT licenses, loadable in one line via SparseEncoder (m7)

Educational thread

In the release thread, tomaarsen explained sparse embedding models: dimensions typically equal vocabulary size, about 99% are zeros, with only a few active dimensions (SPLADE-style). Sparsity yields cheap storage (only non-zero entries are stored) and interpretability, since active dimensions map to concrete tokens (m3, m4, m5).

Why it matters

• Open weights, small size, and leading BEIR scores lower the barrier to production sparse retrieval
• First-class Sentence Transformers SparseEncoder support signals strong community adoption
• Sparse retrieval offers unique storage and interpretability advantages over dense vectors; however, the claim that open sparse models are scarce drew pushback, as options like SpladePP already exist

2026-09-17 ~ 2026-09-18 · 12 related posts

• [source] Linkup open-sources SPARSEUP: sub-150M sparse retriever hits 56+ nDCG@10 on BEIR-13, Apache 2.0 — antoine_chaffin · 2026-09-17
• Linkup Open-Weights SparseUp: Strongest Sub-150M Sparse Retriever on BEIR — tomaarsen · 2026-09-17
• What Are Sparse Embeddings: Vocab-Sized Vectors That Are 99% Zero — tomaarsen · 2026-09-17
• Sparse Embeddings Explained: 99% Zeros, Cheap to Store, Human-Readable — tomaarsen · 2026-09-17
• [source] Sentence Transformers SparseEncoder: Sparse Retrieval in a Few Lines — tomaarsen · 2026-09-17
• SPLADE author counters claim: open sparse models with 4.5M downloads already integrated in chromaDB, Qdrant — prithivida · 2026-09-17
• Open sparse embedding models remain rare; 2-year-old releases hit 4.5M downloads — antoine_chaffin · 2026-09-17
• SBERT docs index open sparse encoder models, plus MIT-licensed splade-mini and splade-tiny — tomaarsen · 2026-09-17
• Linkup Open-Sources SPARSEUP: <150M Sparse Retrieval Model Scoring 56+ on BEIR-13 — antoine_chaffin · 2026-09-17
• Linkup open-sources SPARSEUP, a <150M sparse retriever scoring 56+ on BEIR-13 — antoine_chaffin · 2026-09-17
• Linkup open-sources SPARSEUP, strongest public sparse retriever under 150M, beating SparseEncoder models by a wide margin — antoine_chaffin · 2026-09-18

1 near-duplicate retellings: tomaarsen