LightOn Open-Sources SOTA Sparse Retrievers SlideOn and mSlideOn

On October 19, LightOn released two fully open-source SOTA sparse retrievers: the English SlideOn (174M parameters) and the multilingual mSlideOn (339M parameters), with the models and their accompanying sparse autoencoders (SAE) now available on Hugging Face.

Confirmed

SlideOn scores 56.68 on BEIR-15, a SOTA among sparse baselines, edging out the company's own dense retriever DenseOn at 56.20.
mSlideOn, released by the paulomouraj team and trained on the same data with the same distillation setup, clearly beats the dense mDenseOn on most tasks and is SOTA or competitive against both dense and sparse baselines.
Training approach: SAE pretraining followed by supervised fine-tuning, starting from DenseOn/mLateOn to learn efficient sparse representations compatible with inverted indexes; the team also hit the search efficiency frontier via query vector pruning (q-16 to q-full).
Coverage spans English, multilingual, long-context, and code retrieval; SlideOn is positioned as a first-stage retriever, with the official recommendation to stack a multi-vector reranker and a decision model on top to form a cascaded retrieval pipeline.

Why it matters

This is a strong validation of the sparse retrieval route over dense: a smaller, inverted-index-compatible sparse model overtakes dense baselines on standard benchmarks, and full open-sourcing (including the SAE) lowers the barrier to reproduction and deployment — directly useful for teams building retrieval pipelines (RAG, search).

2026-10-19 ~ 2026-10-20 · 7 related posts

Primary sources

LightOn open-sources SlideOn sparse retrievers: 174M/339M params, tops BEIR-15 at 56.68 — IgorCarron · 2026-10-19 23:15 (UTC+08:00)
SlideOn hits sparse SOTA on BEIR-15 at 56.68, beating its dense baseline at 56.20 — IgorCarron · 2026-10-19 23:14 (UTC+08:00)
mSlideOn: multilingual sparse retriever beats its dense counterpart on most tasks — IgorCarron · 2026-10-19 23:14 (UTC+08:00)

LightOn ships SlideOn and mSlideOn sparse retrievers for first-stage retrieval — IgorCarron · 2026-10-19
LightOn open-sources SOTA sparse retrievers SlideOn and mSlideOn at 174M/339M — IgorCarron · 2026-10-19
[source] SlideOn hits sparse SOTA on BEIR-15 at 56.68, beating its dense baseline at 56.20 — IgorCarron · 2026-10-19
[source] mSlideOn: multilingual sparse retriever beats its dense counterpart on most tasks — IgorCarron · 2026-10-19

3 near-duplicate retellings: IgorCarron · IgorCarron · IgorCarron