Linkup Research has released SPARSEUP, an open-source sparse embedding model built on a 149M-parameter ModernBERT backbone. The company reports a score of 56.4 nDCG@10 on the BEIR-13 benchmark, which it describes as the best result it knows of for a public sparse encoder under 150M parameters.
The model uses a logit-based mechanism, though the source provides little additional technical detail. As a sparse embedding model, SPARSEUP is designed for retrieval tasks where interpretable, term-level representations can be advantageous over dense vectors.
Because there is only one source, its claims about benchmark leadership and model design cannot be independently verified. The release is notable mainly for pushing sparse retrieval performance in a compact, open-weight package.