{"version":1,"type":"story","url":"https://digestai.news/story/linkup-research-releases-sparseup-149m-parameter-sparse-embedding-mode","json":"https://digestai.news/story/linkup-research-releases-sparseup-149m-parameter-sparse-embedding-mode.json","markdown":"https://digestai.news/story/linkup-research-releases-sparseup-149m-parameter-sparse-embedding-mode.md","slug":"linkup-research-releases-sparseup-149m-parameter-sparse-embedding-mode","headline":"Linkup Research releases sparseup 149M-parameter sparse embedding model","summary":"Linkup Research announced SPARSEUP, an open‑source sparse embedding model built on a 149 million‑parameter ModernBERT backbone. The model is released under the Apache 2.0 license on Hugging Face and can be loaded via Transformers or Sentence‑Transformers with the appropriate flag. It incorporates three engineering fixes – a logit shift of 15, per‑position top‑12 token expansion, and case folding – to keep vectors sparse and human‑readable.\n\nSPARSEUP scores 56.4 nDCG@10 on the BEIR‑13 benchmark, which Linkup claims is the strongest public vocabulary‑based sparse encoder under 150 M parameters. In the same controlled setting LateOn scores 58.9 and DenseOn 57.9. On MS MARCO the model attains over 97 % recall in roughly 380 µs per query (single‑threaded), with an average of 47 non‑zero terms per query and 190 per document. Training fits on a single H100 GPU and uses LightOn’s fine‑tuning mixture with contrastive learning.","keyPoints":["SPARSEUP is a 149M‑parameter sparse encoder released under Apache 2.0 on Hugging Face.","It achieves 56.4 nDCG@10 on BEIR‑13, the highest among public sparse models under 150M parameters.","On MS MARCO it reaches >97% recall in ~380 µs per query with 47 and 190 non‑zero terms."],"whyItMatters":"An open, efficient sparse encoder lets developers build fast, interpretable retrieval systems without large dense vectors, lowering compute costs and latency for search‑heavy applications.","category":{"slug":"models","name":"Generative AI & Models","url":"https://digestai.news/category/models"},"entities":{"companies":["Linkup Research","LightOn","Hugging Face"],"models":["SPARSEUP","ModernBERT","LateOn","DenseOn","SPLADE‑v3"],"people":["Asif Razzaq"]},"firstPublishedAt":"2026-09-19T07:48:38Z","updatedAt":"2026-09-19T07:48:38Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"MarkTechPost","title":"Linkup Research Releases SPARSEUP: A 149M-Parameter Open-Source Sparse Embedding Model","url":"https://marktechpost.com/2026/09/19/linkup-research-releases-sparseup","publishedAt":"2026-09-19T07:48:38Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[],"thread":null,"cite":{"text":"Digest AI, \"Linkup Research releases sparseup 149M-parameter sparse embedding model\", 19 September 2026, https://digestai.news/story/linkup-research-releases-sparseup-149m-parameter-sparse-embedding-mode","publisher":"Digest AI","title":"Linkup Research releases sparseup 149M-parameter sparse embedding model","datePublished":"2026-09-19T07:48:38Z","url":"https://digestai.news/story/linkup-research-releases-sparseup-149m-parameter-sparse-embedding-mode"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}