Linkup Releases SPARSEUP, an Apache 2.0 Sparse Embedding Model Under 150M

Linkup Research published SPARSEUP on September 17, 2026, releasing an Apache 2.0 sparse embedding model built on a ModernBERT backbone and hosted at Linkup-Platform/linkup-sparseup-embed-v1 on Hugging Face. The company blog frames the model as the missing sparse slot beside LightOn’s open DenseOn and LateOn releases, using the same backbone family and fine-tuning mixture so architecture comparisons stay controlled.

SPARSEUP is a SPLADE-style encoder that maps queries and documents to vocabulary-based sparse vectors. The model card documents three training knobs versus a vanilla SPLADE-like head: a logit shift before ReLU, per-position top-k expansion, and case/space vocabulary folding after pooling. Inference helpers include encode_to_dict, terminal render, and attribution highlight utilities, with Sentence Transformers SparseEncoder support.

What shipped

Public artifacts are the September 17 Linkup blog post and the Hugging Face model card, both stating Apache 2.0. Author-reported BEIR-13 averages on the card put SPARSEUP at 56.4 nDCG@10 without MS MARCO among sub-150M vocabulary sparse encoders in their comparison table; controlled DenseOn/LateOn numbers are also listed there. Treat all retrieval scores as author-reported evaluations. The blog notes sub-millisecond MS MARCO latency under Seismic approximate search in their measurements; that figure is likewise author-reported.

Primary sources are Linkup’s September 17 blog post and the Linkup-Platform/linkup-sparseup-embed-v1 Hugging Face model card.