Linkup Research has released SPARSEUP, an open-source sparse embedding model built for vocabulary-based retrieval. With a 149M-parameter ModernBERT backbone, Apache 2.0 licensing, and a 56.4 average nDCG@10 score on BEIR-13, SPARSEUP puts a powerful sparse encoder into a compact public package.
The release arrived on September 19, 2026, marking Linkup Research’s first open model. Its weights are available on Hugging Face under Apache 2.0, and developers can load the model through Transformers or Sentence Transformers with trust_remote_code=True. That combination gives researchers and retrieval teams a direct path to test sparse search without building the model from scratch.
A Compact Sparse Encoder With a Strong Benchmark Score
Most open retrieval models are dense: they turn text into continuous vectors that search systems compare in embedding space. Sparse models take a different route, producing weights over a vocabulary. That structure can connect retrieval results to terms, indexes, and search methods in a more direct way.
Linkup calls SPARSEUP the strongest public vocabulary-based sparse encoder under 150M parameters. On BEIR-13, it reaches 56.4 nDCG@10, ahead of opensearch-neural-sparse-encoding-doc-v3-gte at 54.6, opensearch-neural-sparse-encoding-v1 at 52.44, ModernBERT-VT at 52.4, splade-v3 at 51.7, and granite-embedding-30m-sparse at 50.6.
The wider comparison shows why the model matters, but also where the trade-offs remain. With the same backbone family and fine-tuning data, LateOn scores 58.9 on BEIR-13, DenseOn scores 57.9, and SPARSEUP scores 56.4. That leaves SPARSEUP 1.52 points behind DenseOn and 2.5 points behind LateOn, even though all three models use identical data.
SPARSEUP also sits below LACONIC-1B, which scores 58.7 on BEIR-13. Still, its 149M-parameter size and Apache 2.0 license make the result notable within the public sparse encoder field under 150M parameters.
How SPARSEUP Cuts Its Sparse Search Footprint
Training starts from the LateOn-unsupervised checkpoint and uses contrastive learning with seven hard negatives sampled from a pool of 50, along with in-batch negatives. SPARSEUP does not use cross-encoder distillation, and its training fits on a single H100 GPU.
The team fixed a vanilla SPLADE model with three changes: logit shifting, per-position top-k, and case folding. Case folding reduces the output dimensions from about 50k to about 34k, shrinking the vocabulary space while preserving the sparse design.
- Queries use a maximum length of 128 tokens.
- Documents use a maximum length of 512 tokens.
- SPARSEUP averages 47 non-zero terms per query on MS MARCO.
- SPARSEUP averages 190 non-zero terms per document on MS MARCO.
- SPLADE-v3 averages 25 non-zero terms per query and 170 per document.
For search, SPARSEUP uses approximate Seismic search, while LightOn reports exact search. On MS MARCO, Seismic reaches over 97% recall against exact search in about 380 microseconds per query. That result gives the model a clear performance target: keep sparse representations efficient enough for fast retrieval without losing most of the exact-search matches.
Where the Model Wins—and Where It Trails
SPARSEUP wins on ArguAna and Touché, and it beats DenseOn on HotpotQA. DenseOn leads on FiQA and DBPedia, showing that the same training setup can produce different strengths across retrieval tasks.
The benchmark gap also changes after decontamination. On decontaminated BEIR, the gap to DenseOn shrinks to 0.17 points. That result places attention on evaluation design as much as on the headline score, especially when models share a backbone family and fine-tuning data.
A second research article published on 17 September 2026 discusses the robustness of sparse neural networks with adaptive topology. Together, the two September 2026 articles point toward a broader research direction: sparse systems must balance benchmark quality, search speed, representation size, and resilience as their network structure changes.
The acknowledgements discuss Yingtao Zhang and Carlo Vittorio Cannistraci. The work lists the European Union project RRF-2.3.1-21-2022-00004, the Air Force Office of Scientific Research, and support from the National Science Foundation through the AccelNet-MultiNet program.
Asif Razzaq is the CEO of Marktechpost AI Media Inc. He launched Marktechpost, an Artificial Intelligence Media Platform that has over 2 million monthly views.
SPARSEUP now gives the open retrieval community a compact sparse encoder, a reproducible Apache 2.0 release, and clear measurements for judging the next generation of search models. The next challenge is already visible: push sparse retrieval closer to the strongest dense systems while keeping its term-level structure, small footprint, and fast Seismic search.
Based on
