Linkup introduced a sparse search model called Sparseup-embed-v1

Company Linkup released linkup-sparseup-embed-v1 — a specialized model for sparse retrieval, built on the ModernBERT architecture. The solution is distributed under the Apache 2.0 license and is designed for high-performance search systems.

The model demonstrates a result of 56.4 nDCG@10 on the BEIR-13 benchmark. A key feature of the implementation is the low inference latency, which is less than 1 ms per query. Such performance is achieved due to the optimization of the architecture for fast search tasks on large data arrays.

The model's toolkit includes specific token processing methods, such as logit_shift and vocab folding. These functions allow for efficient compression of vector representations and acceleration of the query matching process with documents in indexes.

Sparse retrieval methods remain critically important for scalable RAG systems, where high accuracy is required with limited computational resources. The use of ModernBERT as a basis provides the model with better compatibility with modern natural language processing pipelines compared to classical approaches based on BERT.