Linkup Open-Sources SparseUp, Top Sub-150M Sparse Retrieval Model

Linkup Research (under LightOn) has open-sourced SparseUp, a sparse embedding model under Apache 2.0, filling in the missing sparse retrieval piece alongside DenseOn and LateOn. The model has fewer than 150M parameters, built on ModernBERT (roughly 140M-parameter scale), and achieves 56+ nDCG@10 on the BEIR-13 benchmark — described as the strongest known sparse retrieval model at that size.

Confirmed

• Model released by Linkup Research with open weights, Apache 2.0 license (m1, m2)
• Parameters <150M, backbone is ModernBERT (m1, m2)
• 56+ nDCG@10 on BEIR-13, the strongest known at this scale (m1, m2)
• Model name is linkup-sparseup-embed-v1 (m5)
• Sentence Transformers maintainer tomaarsen gave it a high-profile endorsement and shipped accompanying support (m2, m5)

Primer and Ecosystem Support

In the launch thread, tomaarsen offered a primer on sparse embedding models: vector dimensions typically equal the vocabulary size, with about 99% of positions set to 0 and only a few active dimensions (SPLADE-style methods). Sparsity brings two key benefits: storage only needs to hold non-zero entries, making it cheap, and active dimensions map to specific tokens, making the model highly interpretable (m3, m4).

On the ecosystem side, Sentence Transformers added a new SparseEncoder interface, fully compatible with SparseUp, so sparse embeddings can be up and running in a few lines of code (m5).

Why It Matters

• Open-source + small parameter count + leading BEIR scores lower the barrier to adopting production-grade sparse retrieval
• First-class support from a mainstream framework (Sentence Transformers) signals strong community recognition
• Sparse retrieval offers unique advantages over dense vectors in storage cost and interpretability, yet open-source options have been scarce until now

2026-09-17 ~ 2026-09-17 · 6 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
• [source] 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
• Linkup Open-Sources SPARSEUP: <150M Sparse Retrieval Model Scoring 56+ on BEIR-13 — antoine_chaffin · 2026-09-17