You can deploy this model and start earning money today!

lightonai/
mLateOn

Name: mLateOn
Availability: InStock

sentence-transformers

Views

No views yet

Website

📚 Collection | 📝 Multilingual Blog | 📝 English Blog | 📝 Paper

mLateOn

State-of-the-Art Multilingual ColBERT Retrieval Model by LightOn

🎯 TL;DR: A 307M-parameter multilingual ColBERT (multi-vector) retrieval model achieving state-of-the-art results across multilingual retrieval (MIRACL), long-document retrieval (MLDR), English general-domain retrieval (BEIR), and code retrieval (MTEB Code). Trained on nine languages only, yet generalizes to unseen languages and scripts, demonstrating that late interaction enables translate-train viability for broad multilingual coverage without exhaustive translation.

About the mDenseOn / mLateOn Family

With DenseOn and LateOn, we demonstrated that an open, carefully curated data recipe can match closed-data retrieval models on English. mDenseOn and mLateOn extend this recipe to multilingual, long-context, and code retrieval.

Rather than independently collecting multilingual corpora from scratch (which would be expensive, uneven across languages, and hard to curate at the same quality), we applied the translate-train approach: machine-translating our validated English data into eight target languages (French, German, Italian, Spanish, Portuguese, Swedish, Norwegian, and Arabic) and adding cross-lingual pairs for cross-lingual alignment.

For more information, please read our multilingual models blog post, our English models blog post and our paper.

mLateOn

mLateOn is a multilingual ColBERT (multi-vector) retrieval model built on mmBERT-base (307M parameters), trained by LightOn using PyLate. It supports context lengths of up to 8,192 tokens for documents and queries, using MaxSim scoring.

mLateOn notably:

Achieves the highest BEIR score among all evaluated models at 57.56 NDCG@10, surpassing even the English-only LateOn (57.22) and dense baselines up to twice its size, showing that multilingual training can actually improve English performance.
Leads all models on MIRACL target languages at 65.61, and remains competitive on the full benchmark (67.04) despite being trained on only nine languages, outperforming most models trained on many more languages.
Opens a massive gap on MLDR at 87.69 on target languages, more than 9 points ahead of the next-best model (LFM2.5-ColBERT-350M at 78.39), combining the multilingual generalization of late interaction with ColBERT's affinity for long-context retrieval.
Generalizes to unseen languages, including scripts not seen during retrieval training (Cyrillic, Japanese, etc), demonstrating that multi-vector representations can transfer to languages absent from retrieval training.
Achieves strong code retrieval at 73.48 on MTEB Code, using only fine-tuning-stage code data.

Perhaps the most important finding: the translate-train recipe applied to a dense encoder produces limited generalization beyond target languages, but applying it to a late-interaction architecture lifts this restriction. With mLateOn, you don't need to exhaustively translate to all target languages, a handful of representative languages is sufficient for broad multilingual coverage.

Alongside mLateOn, we also trained mDenseOn, a dense (single-vector) variant using the same setup. If your setup does not allow multi-vector, mDenseOn is a strong single-vector alternative, but trades off generalization to unseen languages and long-context performance.

See our multilingual blog post for full results and analysis.

Results

Headline Results (NDCG@10)

MIRACL_tgt and MLDR_tgt include only the languages overlapping with our target languages; MIRACL and MLDR include all benchmark languages.

Model	Size	BEIR	MIRACL_tgt	MIRACL	MLDR_tgt	MLDR	Code
Our models
mLateOn	307M	57.56	65.61	67.04	87.69	77.92	73.48
mDenseOn	307M	56.70	59.61	58.02	64.98	51.59	71.53
Dense baselines
pplx-embed-v1-0.6b	596M	56.70	63.18	68.20	57.08	43.98	75.18
jina-v5-text-small	677M	56.70	61.03	66.56	53.00	43.83	73.01
jina-v5-text-nano	239M	56.08	60.86	65.84	56.71	47.20	70.75
arctic-embed-l-v2	568M	55.55	60.78	66.53	57.01	48.03	53.17
Qwen3-Embedding-0.6B	596M	55.33	57.64	60.62	59.11	50.06	73.75
harrier-oss-v1-0.6b	596M	55.04	57.41	63.68	53.84	44.23	70.97
harrier-oss-v1-270m	268M	50.82	48.95	57.11	52.31	42.63	63.79
embeddinggemma-300m	308M	53.69	58.72	64.58	49.11	40.89	68.81
gte-multilingual-base	305M	51.08	57.39	64.14	66.41	56.65	57.46
voyage-4-nano	340M	49.96	49.29	58.65	62.10	51.97	76.43
BGE-M3	568M	48.78	62.22	69.62	61.16	52.47	51.49
granite-311m-r2	312M	49.24	53.38	59.74	50.53	41.86	63.50
Late-interaction baselines
pplx-embed-v1-late-0.6b	596M	56.11	63.67	69.55	68.69	55.51	63.29
LFM2.5-ColBERT-350M	353M	54.50	57.40	42.45	78.39	61.05	51.62
jina-colbert-v2	559M	52.96	62.19	65.65	13.97	11.54	49.88
GTE-ModernColBERT	149M	54.23	43.96	36.35	67.68	45.95	54.37
ColBERT-Zero	149M	55.82	43.89	39.36	71.32	49.50	53.59

Key Findings

English quality improved. mLateOn reaches 57.56 on BEIR, higher than the English-only LateOn (57.22). Multilingual training does not degrade English performance; it can enhance it.

Strongest multilingual retrieval on target languages. mLateOn leads all models at 65.61 on MIRACL target languages, ahead of pplx-embed-v1-late-0.6b (63.67) and bge-m3 (62.22), both roughly twice its size and trained on substantially more languages.

Exceptional long-document retrieval. On MLDR target languages, mLateOn achieves 87.69, a gap of around 9 points over the next-best model. This advantage stems from two complementary sources: the stronger multilingual generalization of late interaction, and ColBERT-style models' well-documented affinity for long-context retrieval, where fine-grained token-level matching exploits longer documents more effectively than a single pooled vector.

Late interaction generalizes beyond the retrieval-training languages. On the full MIRACL, mLateOn actually gains from target to full (67.04 vs. 65.61), highlighting exceptional performance on languages never seen during retrieval training. This is in contrast to mDenseOn, which drops about 1.6 points. Even though we trained mostly on Romance languages (plus Arabic), we observe very strong performance on other scripts such as Cyrillic and Japanese.

For practitioners: with a late-interaction model, you don't need to exhaustively translate to all target languages. Translating to a handful of representative languages is sufficient for strong multilingual coverage.

Related Checkpoints

Model	Description	Link
mLateOn (this card)	Multilingual late-interaction retriever (strongest multilingual)	`lightonai/mLateOn`
mLateOn-unsupervised	Multilingual late-interaction, pre-training only	`lightonai/mLateOn-unsupervised`
mDenseOn	Multilingual dense retriever	`lightonai/mDenseOn`
mDenseOn-unsupervised	Multilingual dense, pre-training only	`lightonai/mDenseOn-unsupervised`
LateOn	English-only late-interaction retriever	`lightonai/LateOn`
LateOn-unsupervised	English-only late-interaction, pre-training only	`lightonai/LateOn-unsupervised`
DenseOn	English-only dense retriever	`lightonai/DenseOn`
DenseOn-unsupervised	English-only dense, pre-training only	`lightonai/DenseOn-unsupervised`

Training Data

We release all datasets used to train our English and multilingual embedding models.

Pre-training Data

Multilingual pre-training curated data — 2.16B pairs
English pre-training curated data — 665M pairs

Together these form the 2.8B-pair ready-to-train pre-training mixture. The multilingual corpus spans the eight target languages and includes a 220M-pair cross-lingual split. The curated English corpus is built after filtering and deduplicating the raw English data. The complete 1.4B-pair annotated parent dataset is released as English pre-training unfiltered data.

Fine-tuning Data

A ready-to-train fine-tuning mixture of 16.3M samples, filtered with NV-Retriever methodology to remove false negatives: any mined candidate scoring above 95% of the positive's relevance is discarded, and the top 10 surviving hard negatives are kept per sample. The mixture is then annotated with the mxbai-rerank-large-v2 cross-encoder for distillation. We release one dataset per language, plus two code splits:

Multilingual data. MIRACL and MLDR hard-negatives were mined directly in the available datasets with snowflake-arctic-embed-l-v2.0 (2,048 mined negatives per sample) and the full unfiltered set is available at embeddings-fine-tuning-multilingual-unfiltered. All other splits follow the translate-train recipe: we apply the mentioned NV-Retriever filtering to the English fine-tuning unfiltered data, then machine-translate the filtered samples into the eight target languages, carrying the original relevance scores. The unfiltered English fine-tuning parent corpus contains 1.88M examples with 2,048 mined negatives, scored with GTE-ModernBERT.

Code data. The Code dataset is derived from LateOn-Code data, where 2,048 negatives are mined with GTE-ModernBERT. The Code Edit dataset had no existing training set, so we built one from CommitPackFT, decontaminated against the CodeEditSearch evaluation set by removing shared commit SHAs, exact normalized-text matches, and examples with ≥50% 13-gram containment overlap. Hard-negatives are likewise mined with GTE-ModernBERT, and its unfiltered mined set is released alongside the multilingual data in embeddings-fine-tuning-multilingual-unfiltered.

For more information, please read our multilingual blog post and our English blog post. Training scripts can be found here.

Model Details

Model Description

Model Type: PyLate ColBERT model
Base model: mmBERT-base
Query/Document Length: 8,192 tokens
Output Dimensionality: 128 dimensions
Similarity Function: MaxSim
Languages: English, French, German, Italian, Spanish, Portuguese, Swedish, Norwegian, Arabic
License: Apache 2.0

Model Sources

Documentation: PyLate Documentation
Repository: PyLate on GitHub
Hugging Face: PyLate models on Hugging Face

Full Model Architecture

ColBERT(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False, 'architecture': 'ModernBertModel'})
  (1): Dense({'in_features': 768, 'out_features': 1536, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity', 'use_residual': True})
  (2): Dense({'in_features': 1536, 'out_features': 768, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity', 'use_residual': True})
  (3): Dense({'in_features': 768, 'out_features': 128, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity', 'use_residual': False})
)

Usage

Sentence Transformers

This model can be used with Sentence Transformers as a multi-vector (ColBERT-style late interaction) retriever via the MultiVectorEncoder:

pip install "sentence-transformers>=6.0.0"

python

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("lightonai/mLateOn")

query = "Which planet is the Red Planet?"
documents = [
    "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
    "Mars, connu pour son apparence rougeâtre, est souvent appelé la planète rouge.",
    "Mars, bekannt für sein rötliches Erscheinungsbild, wird oft als der Rote Planet bezeichnet.",
    "Venus is often called Earth's twin because of its similar size and proximity.",
]

query_embeddings = model.encode_query(query)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings[0].shape)
# (10, 128) (20, 128)

# MaxSim late-interaction scoring (higher is more relevant)
scores = model.similarity(query_embeddings, document_embeddings)
print(scores)
# tensor([[9.6029, 9.5838, 9.5877, 9.4578]])

PyLate

First install the PyLate library:

pip install -U pylate

Retrieval

Use this model with PyLate to index and retrieve documents. The index uses FastPlaid for efficient similarity search.

Indexing documents

Load the ColBERT model and initialize the PLAID index, then encode and index your documents:

python

from pylate import indexes, models, retrieve

# Step 1: Load the ColBERT model
model = models.ColBERT(
    model_name_or_path="lightonai/mLateOn",
)

# Step 2: Initialize the PLAID index
index = indexes.PLAID(
    index_folder="pylate-index",
    index_name="index",
    override=True,  # This overwrites the existing index if any
)

# Step 3: Encode the documents (supports multilingual content)
documents_ids = ["1", "2", "3", "4"]
documents = [
    "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
    "Mars, connu pour son apparence rougeâtre, est souvent appelé la planète rouge.",
    "Mars, bekannt für sein rötliches Erscheinungsbild, wird oft als der Rote Planet bezeichnet.",
    "Venus is often called Earth's twin because of its similar size and proximity.",
]

documents_embeddings = model.encode(
    documents,
    batch_size=32,
    is_query=False,  # Ensure that it is set to False to indicate that these are documents, not queries
    show_progress_bar=True,
)

# Step 4: Add document embeddings to the index by providing embeddings and corresponding ids
index.add_documents(
    documents_ids=documents_ids,
    documents_embeddings=documents_embeddings,
)

Note that you do not have to recreate the index and encode the documents every time. Once you have created an index and added the documents, you can re-use the index later by loading it:

python

# To load an index, simply instantiate it with the correct folder/name and without overriding it
index = indexes.PLAID(
    index_folder="pylate-index",
    index_name="index",
)

Retrieving top-k documents for queries

Once the documents are indexed, you can retrieve the top-k most relevant documents for a given set of queries — including cross-lingual retrieval:

python

# Step 1: Initialize the ColBERT retriever
retriever = retrieve.ColBERT(index=index)

# Step 2: Encode the queries (can be in any supported language)
queries_embeddings = model.encode(
    [
        "Quelle planète est connue comme la planète rouge ?",  # French query
        "Which planet is the Red Planet?",                      # English query
    ],
    batch_size=32,
    is_query=True,  # Ensure that it is set to True to indicate that these are queries
    show_progress_bar=True,
)

# Step 3: Retrieve top-k documents (returns matches across all languages)
scores = retriever.retrieve(
    queries_embeddings=queries_embeddings,
    k=10,  # Retrieve the top 10 matches for each query
)

Reranking

If you only want to use the ColBERT model to perform reranking on top of your first-stage retrieval pipeline without building an index, you can simply use the rank function and pass the queries and documents to rerank:

python

from pylate import rank, models

queries = [
    "Quelle planète est la planète rouge ?",
    "What is the largest planet in the solar system?",
]

documents = [
    ["Mars ist der Rote Planet.", "Venus ist der Zwilling der Erde."],
    ["Jupiter is the largest planet.", "Saturn is known for its rings.", "Mars is the Red Planet."],
]

documents_ids = [
    [1, 2],
    [1, 3, 2],
]

model = models.ColBERT(
    model_name_or_path="lightonai/mLateOn",
)

queries_embeddings = model.encode(
    queries,
    is_query=True,
)

documents_embeddings = model.encode(
    documents,
    is_query=False,
)

reranked_documents = rank.rerank(
    documents_ids=documents_ids,
    queries_embeddings=queries_embeddings,
    documents_embeddings=documents_embeddings,
)

Framework Versions

Python: 3.11.10
Sentence Transformers: 5.1.1
PyLate: 1.3.4
Transformers: 4.57.5
PyTorch: 2.9.0+cu128
Accelerate: 1.12.0
Datasets: 3.6.0
Tokenizers: 0.22.1

Citation

BibTeX

mDenseOn and mLateOn

bibtex

@misc{sourty2026denseonlateonfullyopen,
  title         = {DenseOn with the LateOn: Fully Open Dense and Late-Interaction Models for Multilingual, Long-Context, and Code Search},
  author        = {Raphaël Sourty and Antoine Chaffin and Paulo Roberto Moura Junior and Amélie Chatelain},
  year          = {2026},
  eprint        = {2607.27178},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2607.27178},
}

DenseOn and LateOn

bibtex

@misc{sourty2026denseonlateon,
  title={DenseOn with the LateOn: Open State-of-the-Art Single and Multi-Vector Models},
  author={Sourty, Raphael and Chaffin, Antoine and Weller, Orion and Moura Junior, Paulo Roberto and Chatelain, Amelie},
  year={2026},
  howpublished={\url{https://huggingface.co/blog/lightonai/denseon-lateon}},
}

PyLate

bibtex

@inproceedings{DBLP:conf/cikm/ChaffinS25,
  author       = {Antoine Chaffin and
                  Rapha{\"{e}}l Sourty},
  editor       = {Meeyoung Cha and
                  Chanyoung Park and
                  Noseong Park and
                  Carl Yang and
                  Senjuti Basu Roy and
                  Jessie Li and
                  Jaap Kamps and
                  Kijung Shin and
                  Bryan Hooi and
                  Lifang He},
  title        = {PyLate: Flexible Training and Retrieval for Late Interaction Models},
  booktitle    = {Proceedings of the 34th {ACM} International Conference on Information
                  and Knowledge Management, {CIKM} 2025, Seoul, Republic of Korea, November
                  10-14, 2025},
  pages        = {6334--6339},
  publisher    = {{ACM}},
  year         = {2025},
  url          = {https://github.com/lightonai/pylate},
  doi          = {10.1145/3746252.3761608},
}

Sentence Transformers

bibtex

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084"
}

Acknowledgements

We thank Eugene Yang for his feedback on adapting our English study to multilinguality through translate-train. We again thank Xin Zhang, Zach Nussbaum, Tom Aarsen, Bo Wang, Eugene Yang, Benjamin Clavié, Nandan Thakur, Oskar Hallström and Iacopo Poli for their valuable contributions and feedback on the original English study. We thank Orion Weller for building the FineWeb-derived Common Crawl split as well as for his feedback and help. We are grateful to the teams behind Sentence Transformers, BEIR, and MIRACL, and to the open-source retrieval community, in particular the authors of Nomic Embed.

This work was granted access to the HPC resources of IDRIS under GENCI allocations AS011016449, A0181016214, and A0171015706 (Jean Zay supercomputer). We also acknowledge the Barcelona Supercomputing Center (BSC-CNS) for providing access to MareNostrum 5 under EuroHPC AI Factory Fast Lane project EHPC-AIF-2025FL01-445. This project is also supported by the OpenEuroLLM project, co-funded by the Digital Europe Programme under GA no. 101195233.