Linkup's SPARSEUP Model Tops BEIR-13 Benchmark

Linkup's research team put out SPARSEUP this week, a learned sparse embedding model built on a 149M-parameter ModernBERT backbone and released under Apache 2.0. On the BEIR-13 benchmark, the standard test suite for retrieval quality, the model hit 56.4 average nDCG@10, which Linkup says makes it the strongest public vocabulary-based sparse encoder under 150M parameters that the team is aware of. Weights are already up on Hugging Face, and the model loads through Transformers or Sentence Transformers with trust_remote_code=True, so it's usable today rather than a research curiosity.

The release exists because most open retrieval models are dense, outputting one vector per piece of text. Sparse models take a different route, assigning weights across a vocabulary so each dimension corresponds to an actual token. That makes the output readable and lets it slot into inverted indexes, plus it tends to handle rare words better than dense approaches.

Linkup built SPARSEUP after LightOn released DenseOn and LateOn, a dense model and a late-interaction model trained on shared data and a shared backbone. Sparse was the missing piece, and how Linkup got there involved fixing a specific problem: a default SPLADE setup on this backbone kept filling result sets with stopwords.