Linkup opens a 149M sparse retriever

Linkup Research released SPARSEUP, a 149-million-parameter open sparse embedding model under Apache 2.0. In Linkup’s release, it reports 56.4 average nDCG@10 on BEIR-13 and about 380 microseconds per query on MS MARCO at more than 97% recall against exact search; it also says the model trails comparable dense and late-interaction systems, so sparse is not a universal replacement. Retrieval builders should benchmark it where lexical traceability, inverted-index latency, or Apache licensing matter, alongside stateless MCP infrastructure that scales like ordinary HTTP rather than assuming it wins semantic search.