HuggingFace released SlideOn and mSlideOn on October 19, 2026, saying it provides state-of-the-art sparse retrieval for English and multilingual tasks. It is the company's first sparse retrieval update since the release of DenseOn and LateOn.
HuggingFace reported 56.68 on BEIR-15, measured on the BEIR-15 benchmark. That compares with DenseOn’s 56.20.
SlideOn and mSlideOn are built on SAE pretraining and supervised fine-tuning, targeting efficient inverted-index search for large document collections. Availability begins with open-source release, initially for researchers and developers.
"Through SAE pretraining and supervised fine-tuning, we build SlideOn from DenseOn and mSlideOn from mLateOn, reusing their existing retrieval knowledge to learn sparse representations compatible with efficient inverted-index search," said Raphael Sourty, Lead Researcher. The models maintain strong retrieval performance as pruning increases, allowing users to adjust them to fit their latency and storage budgets.
The announcement follows the release of DenseOn and LateOn, which validated the data recipe and training setup. HuggingFace said the results encouraged them to build on previous work and investigate how to convert dense and late-interaction models into efficient state-of-the-art sparse retrievers.
HuggingFace did not say how the models will perform on other benchmarks, and raised the open question of how sparse retrieval will scale to even larger document collections. The company said it will continue to release SAE checkpoints, fine-tuned retrieval models, and training code under a permissive license to support further research.
Source: huggingface
