LightOn open-sources DenseOn: SOTA single and multi-vector retrieval models
antoine_chaffin · x · 2026-11-13
LightOn released the DenseOn family on Hugging Face: open single-vector (dense) and multi-vector (ColBERT-style late interaction) retrieval models that hit open-model SOTA on BEIR, with the multi-vector models showing strong multilingual generalization of late interaction architectures.
The article fully documents the pipeline: pre-training data mixture and filtering, fine-tuning data construction, and results on BEIR including a decontaminated split.
Author Antoine Chaffin stresses the released data's value beyond the models — it already enabled Ettin rerankers, the MICE extension, SparseUp and more: "open source creates a race to the top." Teams still training only on MSMARCO are urged to check the English/multilingual blog posts and paper.
More from Models
- Leaderboards miss vision skills: why DeepSeek's chart-reading progress doesn't show up in rankings — TwoBlueFools · 2026-11-13
- OpenAI's 1147-page Yang-Mills proof faces its first expert assessment: not proven yet — mishig25 · 2026-11-13
- Terminal-Bench 4.0 Cost Math: GPT-6.1 Sol Matches Opus 5.5 at ~1/8 the Price — LeadershipNo2904 · 2026-11-13
- HF staffer spotlights basal-2.0-mini, a 1.5B decision-tuned model from Bielik — lhoestq · 2026-11-13
- Observer surprised OpenAI shipped no answer to rival's new model release — jdjohnson · 2026-11-13
- Google's Carbon-Argon-Hydrogen model naming breaks its own size logic — JeremyNguyenPhD · 2026-11-13