Skip to main content
Back

Linkup's SPARSEUP Beats Every Sub-150M Sparse Search Model With a 56.4 Score

Linkup released SPARSEUP, a 150M sparse retriever that beats every sparse model under 1B on BEIR and hits sub-millisecond search latency.

  • Linkup Research released SPARSEUP, a 150M sparse retriever built on ModernBERT, Apache 2.0.
  • Hits 56.4 nDCG@10 on BEIR-13, beating every public sparse encoder under 1B parameters.
  • Reaches sub-millisecond query latency on MS MARCO when paired with the Seismic index.
  • Three new tricks: logit_shift=15, per-token top-k=12 expansion cap, and case/space vocab folding.
  • Co-authored by Thibault Formal, original SPLADE author, and fully compatible with Sentence Transformers.
  • Full details in the Linkup blog post.

Linkup Research has released SPARSEUP, an Apache 2.0, SPLADE-style retriever for English search. Linkup reports the highest BEIR-13 score among public vocabulary-based sparse models below 150M parameters, along with sub-millisecond indexed search in a specific Seismic benchmark. The release gives developers a compact sparse option with learned query expansion, token-level diagnostics, and Sentence Transformers support.

Vocabulary weights you can inspect

SPLADE-style retrievers represent each query and document as a weighted vector over the model’s vocabulary. A vocabulary of roughly 50,000 tokens produces a vector with roughly 50,000 possible dimensions, but most values remain zero. Similarity is computed from the dimensions that the query and document share, typically with a sparse dot product.

Each nonzero coordinate maps to a specific token, which makes the representation easier to inspect than a dense vector of latent dimensions. Learned expansion can also activate relevant terms absent from the source text. Those active dimensions can be stored as weighted token postings, although expansion may enlarge posting lists and increase index size.

A cleaner three-way architecture test

LightOn previously released open data and training recipes for DenseOn, a single-vector dense encoder, and LateOn, a ColBERT-style late-interaction model. SPARSEUP adds learned sparse retrieval to the same comparison series, with backbone and training data held steady. That setup reduces two major confounders when comparing retrieval architectures. Thibault Formal, an author of the original SPLADE work, co-authored the release post.

A 56.4 score with a defined scope

BEIR measures retrieval quality across datasets and domains. Its nDCG@10 metric rewards systems for placing graded relevant results near the top of the first 10 results, with higher scores indicating better ranking quality. Linkup reports the following average across 13 BEIR datasets, excluding MS MARCO: