Linkup Research released SPARSEUP, an open-source learned sparse embedding model running on a 149M-parameter ModernBERT backbone and shipped under Apache 2.0. The team reports 56.4 average nDCG@10 on BEIR-13 and calls it the strongest public vocabulary-based sparse encoder it knows of under 150M parameters. Weights are on Hugging Face, and the model loads through Transformers or Sentence Transformers with trust_remote_code=True. The trigger was LightOn’s DenseOn and LateOn release; SPARSEUP fills the missing sparse slot using the same backbone family and fine-tuning data so all 3 retrieval styles can be compared side by side. Training starts from LateOn-unsupervised, grafting back ModernBERT’s original MLM head, then fine-tuning with LightOn’s fine-tuning mixture and contrastive learning only, with 7 hard negatives sampled from a pool of 50 plus in-batch negatives, no cross-encoder distillation, fitting on a single H100. To fix a vanilla SPLADE’s stopword-heavy bags, Linkup used logit shifting via log(1 + ReLU(x - 15)), per-position top-k keeping each token’s 12 strongest vocabulary dimensions before max pooling, and case folding, dropping output dimensions from about 50k to about 34k. Queries and documents take [Q] and [D] prefixes, scoring is a dot product