DenseOn

📚 Collection | 📝 Blog

🎯 TL;DR: A 149M-parameter dense (single-vector) retrieval model achieving 56.20 nDCG@10 on BEIR, the first sub-150M dense model to break the 56 mark, topping all base-size dense models and outperforming several models 4× larger.

About the LateOn / DenseOn Family

State-of-the-art retrieval is increasingly dominated by closed models, either hidden behind APIs or trained on undisclosed data. This blocks reproducibility, prevents study of possible data leakage, and gatekeeps progress to a handful of private labs. We thus decided to gather and curate a large amount of data and explore various mixtures. We release all the data used in our explorations:

• Gathered pre-training data, 1.4B query-documents pairs alongside annotations used for non-destructive filtering (structural filtering, deduplication, cross-encoder pair relevancy)
• Best pre-training mixture found with already applied filters
• Fine-tuning datasets with query, positive and 2048 mined documents alongside their scores for 1.88M samples.

Based on our findings, we trained LateOn (multi-vector/ColBERT) and DenseOn (single vector/dense) models on a proprietary Apache 2.0-compatible training dataset and release those models as well. Both are built on the ModernBERT backbone at 149M parameters, a size we believe sits at the sweet spot: large enough to handle real-world queries and documents, small enough to serve at high throughput in latency-sensitive production systems. For more information, please read our blogpost.

DenseOn

DenseOn is a dense (single-vector) retrieval model built on ModernBERT (149M parameters), trained by LightOn. It encodes queries and documents independently using cosine similarity with query:/document: prefixes and CLS pooling.

DenseOn achieves 56.75 average NDCG@10 on BEIR (15 datasets) and 57.71 on decontaminated BEIR (12 datasets), topping all base-size dense models and outperforming models up to 4x larger. Notably it:

• Tops all base-size dense models on BEIR, ahead of GTE-ModernBERT (55.19) and on par with much larger Snowflake Arctic Embed L v2 (55.22, 568M) and Qwen3-Embedding-0.6B (55.52).
• Holds up under decontamination: when training-overlap samples are stripped from the BEIR corpora, DenseOn improves to 57.71 nDCG@10 on the 12-dataset decontaminated split.

Alongside DenseOn, we also trained LateOn, a late-interaction variant trained using the same setup. It achieves stronger BEIR results shares the usual benefits of late interaction models (better generalization, long context capabilities, ...). For more information about late interaction models, you can check our previous work on the matter such as ColBERT-Zero and give Pylate a shot (the best entry point in the late interaction ecosystem developped at LightOn).

See our blog post for full results and analysis.

Results

BEIR (15 datasets, NDCG@10)