FrontierSeptember 19, 2026via MarkTechPost
Linkup Research Releases SPARSEUP: A 149M-Parameter Open-Source Sparse Embedding Model
Why it matters
A practical open-weight embedding model that achieves state-of-the-art sparse retrieval performance at a size practitioners can run locally or on modest infrastructure. Relevant to builders optimizing RAG, semantic search, and retrieval systems where sparse vectors offer speed and interpretability advantages over dense alternatives.
Key signals
- SPARSEUP: 149M-parameter ModernBERT backbone
- 56.4 nDCG@10 on BEIR-13 benchmark (claimed best-in-class for sub-150M sparse encoders)
- 97% recall achieved in ~380 microseconds per query with Seismic index
- Techniques: logit shift, top-12 token expansion, case folding for sparsity
- Licensed under Apache 2.0 (open-source)
- Released by Linkup Research
- 149M parameters (ModernBERT backbone)
- 56.4 nDCG@10 on BEIR-13 benchmark
- Claims best public result for sparse encoder under 150M parameters
- 97%+ recall in ~380 microseconds per query with Seismic index
- Apache 2.0 license (open-weight)
- Uses logit shift, top-12 expansion per token, case folding for sparsity
The hook
149M parameters. 56.4 nDCG@10. Linkup just shipped the best open sparse encoder under 150M — and it's fast enough for production search.
Linkup Research has released SPARSEUP, an open-source sparse embedding model built on a 149M-parameter ModernBERT backbone. It scores 56.4 nDCG@10 on BEIR-13, which Linkup calls the best result it knows of for a public sparse encoder under 150M parameters. The model uses a logit shift, top-12 expans…