mSlideOn

mSlideOn

Turns queries and documents into sparse vectors you can search with an inverted index, across several languages and long text. A 339M multilingual retrieval model for people building search.

Made by lightonai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: mSlideOn is a multilingual retrieval model from LightOn. It turns each query or document into a sparse vector, meaning most of its 32,768 positions stay empty. That makes it work with ordinary inverted-index search. It started from the mLateOn model and has 339M parameters.

Who it suits: Developers and researchers who build search over text in several languages, long documents or code, and who want sparse vectors rather than dense ones. The record lists English, French, German, Italian, Spanish, Portuguese, Swedish and Norwegian.

Industry | Who uses it | How they use it | Example
Search engineering | Search teams | Add semantic matching to an inverted-index search stack | Matching a Spanish query to French documents
Software | Developers of knowledge tools | Retrieve passages from long documents | Finding the right section of a long report
Code tools | Assistant builders | Retrieve code for a question | Looking up a function from a description
Research | Retrieval researchers | Compare sparse, dense and late-interaction models | Checking scores on public retrieval benchmarks

You can skip it if you only search short English text and already have a dense retriever you like. LightOn's SlideOn is the English version.

Where to use it:

• Python projects that use the sentence-transformers library
• Inverted-index search systems that store sparse vectors
• Windows, macOS or Linux with Python installed

Use case | Who | What they get back
Multilingual search | Search engineers | Documents ranked for a query in another language
Long-document retrieval | Developers | Matches from texts up to 8,192 tokens
Code retrieval | Tool builders | Code ranked against a text query

How to set it up:

• Python on Windows, macOS or Linux
• The huggingface_hub package, and sentence-transformers to run the model
• Room for about 1.4 GB of weights in F32

1. Install the Hugging Face client.
2. Download the model files to your computer.
3. Load the model with the sentence-transformers library.

pip install -U huggingface_hub
hf download lightonai/mSlideOn

How to use it: Encode your documents and your queries with the model, store each document's active positions in an inverted index, then score a query by the dot product of the two vectors. The source shows no code, so follow the model card and sentence-transformers' guide for the exact calls.

Tool | What it does
sentence-transformers | Loads the model and encodes text
Inverted index | Stores each document's active positions for fast lookup
Dot product | Scores a query against a document vector

Example: You index a set of French and German articles, then search with a Spanish question. The model turns each text into a sparse vector, and the dot product ranks the articles that best match the question. You then read the top few.

For newcomers and experts: If you are new, picture each text as a short list of learned topic features with weights, and search as matching those lists. A first try is a handful of documents and a query in a second language.

If you know the field: a sparse autoencoder lifts mLateOn's token representations to 768 dimensions, keeps the top 32 activations and applies ReLU, and max pooling over tokens gives one vector. Query and document sparsity can be adjusted separately through Matryoshka training.

Text
mLateOn token representations
Sparse encoder, top 32 activations
Max pooling to one vector
Inverted-index search

Good to know: The authors report that mSlideOn beats its dense and late-interaction relatives on the MTEB code average.

Title: MTEB Code average, nDCG@10
mDenseOn: 71.53 points
mLateOn: 73.48 points
mSlideOn: 74.78 points

• Input length: up to 8,192 tokens
• Sparsity: adjustable for queries and documents on their own
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: model files downloaded to your computer, plus the text you choose to index
• Not for: the source names none, though the record lists only eight languages

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: sentence-transformers
• Parameters: 339M
• Weights: about 1.4 GB in F32; about 170 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 706
• Likes: 16
• License: apache-2.0
• Languages: en, fr, de, it, es, pt, sv, no
• Based on: lightonai/mLateOn
• Tags: sentence-transformers, safetensors, slideon, sparse, sentence-similarity, custom-code, multilingual, code-search, sparse-retrieval, matryoshka, feature-extraction, code

Links

• Model: https://huggingface.co/lightonai/mSlideOn
• Files: https://huggingface.co/lightonai/mSlideOn/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lightonai/mSlideOn

Then follow the usage in its model card. Nothing in this kit runs by itself.