Linkup Research Releases SPARSEUP: A 149M-Parameter Open-Source Sparse Embedding Model

Linkup Research introduced SPARSEUP, an open sparse embedding model based on ModernBERT (149 million parameters) under the Apache 2.0 license. The model scores 56.4 nDCG@10 on BEIR-13 and achieves over 97% recall in search at ~380 µs per query.

• 149 million parameters, Apache 2.0, weights on Hugging Face
• 56.4 nDCG@10 on BEIR-13 — best among public sparse models up to 150M parameters
• Three refinements: logit shift by 15, top-12 token expansion, case folding
• Dimensionality reduced from ~50k to ~34k via case folding