Linkup ships a 149M retrieval encoder under Apache 2.0

single source· 1 articles · confidence: medium · first seen 2026-09-19 07:48 UTC

What this means for you

Worth a look if you self-host retrieval and want something small enough to run on modest hardware — Apache 2.0, 149M parameters, and a reported 380 microseconds per query with the Seismic index. Treat the BEIR figure as a vendor claim: no evaluation date, harness or third-party reproduction is given.

Linkup Research has released SPARSEUP, a sparse embedding model (text becomes vectors that are mostly zeros, so search runs on keyword-style inverted indexes) on a 149M-parameter ModernBERT backbone. It reports 56.4 nDCG@10 on BEIR-13, a retrieval ranking measure, and says it knows of no better public sparse encoder below 150M parameters. Paired with the Seismic index it reaches over 97% recall at roughly 380 microseconds per query, under Apache 2.0. It uses logit shift, top-12 expansion per token and case folding; no evaluation date or harness is given.

Key facts

  • ·SPARSEUP is a sparse embedding model with 149M parameters, built on a ModernBERT backbone. source
  • ·Linkup reports 56.4 nDCG@10 on BEIR-13, with no evaluation date or harness given. source
  • ·Linkup says it knows of no better public sparse encoder under 150M parameters. source
  • ·With the Seismic index, the model reaches over 97% recall at about 380 microseconds per query. source
  • ·The model is released under Apache 2.0 and uses a logit shift, top-12 expansion per token and case folding. source

What the sources say

  • MarkTechPost — Vendor announcement carrying the benchmark, latency, sparsity method and licence details.

Sources

The original reporting. Follow these — they did the work.

← the wire