SlideOn
by lightonai
LightOn's 174M English sparse retriever: 32,768-dimension learned sparse vectors you can serve from an inverted index, with separate budgets for query and document size
lightonai/SlideOnDeploy SlideOn
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
SlideOn is a sparse retrieval model from LightOn, published on Hugging Face on 12 October 2026 under the Apache-2.0 license. It turns a query or a document into a 32,768-dimension vector in which most dimensions are zero, so the vectors can be stored in an inverted index like keyword search while matching on meaning. It is fine-tuned from LightOn's dense model DenseOn through sparse autoencoder pretraining and supervised retrieval training.
The card reports 56.68 average nDCG@10 on 15 BEIR datasets with unpruned vectors, and 56.66 with queries capped at 64 active dimensions and documents at 512. Its dimensions are learned features, so they do not map to words the way a vocabulary-based model such as SPLADE does. A multilingual sibling, mSlideOn, starts from mLateOn.
Architecture
A DenseOn encoder (174M parameters) followed by a learned linear sparse encoder: each normalized token representation keeps its top-32 latent activations after ReLU, and the pooled vector has 32,768 nonnegative dimensions. Scoring is a dot product. Queries and documents are encoded with their own prompts through encode_query and encode_document, and max_active_dims caps how many dimensions each keeps. Inputs run up to 8,192 tokens.
Mixpeek SDK Integration
# Encode documents with SlideOn (q-64 / d-512, the card's example budgets).
# pip install -U "sentence-transformers>=6.1,<6.2" "transformers>=5.3,<6"
from sentence_transformers import SparseEncoder
model = SparseEncoder("lightonai/SlideOn", trust_remote_code=True)
docs = ["A solar eclipse occurs when the Moon passes between the Sun and Earth."]
doc_vecs = model.encode_document(docs, max_active_dims=512) # torch sparse, 32,768 dimsCapabilities
- Learned sparse vectors that fit an inverted index
- Separate size budgets for queries and documents (max_active_dims)
- Inputs up to 8,192 tokens
- Loads with Sentence Transformers SparseEncoder
- Apache-2.0 license
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| BEIR, 15 datasets | Mean nDCG@10 | 56.68 | LightOn model card, unpruned query and document vectors |
| BEIR, 15 datasets | Mean nDCG@10 (queries 64, documents 512 active dims) | 56.66 | LightOn model card |
Performance
The card's retrieval FLOPS proxy drops to 8.72 with queries capped at 64 active dimensions and documents at 512. Sparse vectors need a store that indexes nonzero dimensions; returned embeddings are PyTorch sparse tensors. The dimensions are learned latents, so SparseEncoder.decode() cannot label them as words. We have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is SlideOn?
A 174M-parameter English sparse retrieval model from LightOn, released under Apache-2.0. It encodes text into 32,768-dimension vectors with few nonzero values, which an inverted index can search, and reports 56.68 mean nDCG@10 on 15 BEIR datasets.
How is SlideOn different from SPLADE?
SPLADE's dimensions are vocabulary terms, so each nonzero weight belongs to a word. SlideOn's 32,768 dimensions are features learned by a sparse autoencoder over DenseOn's token representations, and they do not map to words.
What does max_active_dims do?
It keeps the N largest weights in each vector and zeroes the rest. The card's example uses 64 for queries and 512 for documents, which it reports at 56.66 nDCG@10 against 56.68 unpruned.
How do I use SlideOn with Mixpeek?
Mixpeek's managed extractors do not include SlideOn today; running it as an extractor means registering a custom model, an Enterprise path. To compare it with your current search first, encode a sample of your documents as in the example on this page.
Specification
Build a pipeline with SlideOn
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free