What is LateOn and mLateOn?
LateOn is a ColBERT-style late-interaction model from LightOn, the Paris AI company, released on 21 April 2026 under Apache 2.0. Instead of squeezing a document into one vector, it keeps a 128-number vector for every token and scores a query by matching each query token to its closest document token (MaxSim). That costs more storage than a dense embedding but tends to rank more precisely, and the same model can either search a prebuilt index or rerank another system's shortlist.
LateOn has 149 million parameters on ModernBERT and is English-only. mLateOn, published on 30 July 2026, is the multilingual follow-up: 307 million parameters on mmBERT, trained on nine languages (English, French, German, Italian, Spanish, Portuguese, Swedish, Norwegian and Arabic) and documents of up to 8,192 tokens.
LightOn released the training data behind both: 1.4 billion gathered query-document pairs, the curated pre-training mix and the fine-tuning sets with mined negatives. Both run with LightOn's PyLate library and its FastPLAID index.
Is LateOn and mLateOn worth it?
LateOn moved the bar for small retrievers. It averages 57.22 NDCG@10 on the 15 BEIR datasets, ahead of LightOn's own ColBERT-Zero at 55.39 and GTE-ModernColBERT at 54.75, and well ahead of Jina-ColBERT-v2 at 51.85, which is roughly four times its size. On a 12-dataset version of BEIR stripped of passages that overlap with training data it rises to 60.36, which answers the usual suspicion that leaderboard leads come from leakage.
mLateOn matters more for most production corpora. It scores 57.56 on BEIR, slightly above the English model, 65.61 on MIRACL's target languages and 87.69 on MLDR, a long-document benchmark, more than nine points ahead of Liquid AI's LFM2.5-ColBERT-350M on LightOn's figures. LateOn itself is weak outside English: Liquid's multilingual tests put it at 0.484 on NanoBEIR against 0.605 for Liquid's model, while LateOn has the best English score in the same table at 0.690.
The cost of late interaction is storage. Keeping one 128-number vector per token means a 300-token passage stores about 300 vectors where a dense model stores one, so indexes run far larger and need a ColBERT-aware engine such as FastPLAID. Used as a reranker over another system's top 100, nothing is stored and the cost is simply encoding those documents per query.
On reasoning-heavy queries it is an unknown: LightOn publishes no BRIGHT results, and the hosted cross-encoders reviewed here report them. LateOn's default configuration also caps documents at 300 tokens, so long passages need mLateOn or chunking. For teams that want open data, open weights and a permissive licence in one package, this pair is where late interaction starts.
Where it earns its place
An ML engineer at a legal-tech start-up with long French and German contracts
Retrieve the right clause across languages without a paid API
Index contracts with mLateOn in PyLate's FastPLAID index, using its 8,192-token window to keep sections whole. Apache 2.0 means no licence fee; the trade is a larger index than a dense model would need.
A search engineer with BM25 already in Elasticsearch
Rerank the top 100 keyword hits without building a vector index
Encode the query and the 100 candidates with LateOn and score them with PyLate's rank.rerank, which needs no index. Cost is GPU or CPU time for 149M parameters per query.
A research team auditing retrieval models for data leakage
Reproduce or extend a retriever's training on known data
Start from LightOn's released pre-training and fine-tuning sets and the 12-dataset decontaminated BEIR split, then train with PyLate. LightOn reports LateOn's score on the clean split at 60.36.
Before you commit
- Setup
- pip install pylate, load models.ColBERT with lightonai/LateOn or lightonai/mLateOn, encode documents into a PLAID index, then retrieve; reranking a shortlist needs only rank.rerank and no index.
- Integrations
- PyLate and FastPLAID from LightOn, built on Sentence Transformers; weights on Hugging Face. Storing token vectors elsewhere needs a vector database with multi-vector MaxSim support.
- Data and privacy
- Self-hosted only, so documents stay on your own hardware. Training data for the experiments is public on Hugging Face; the final models used a proprietary Apache-compatible dataset.
- Cost at scale
- No licence or API fee. The running cost is GPU time for encoding and storage for one 128-dimension vector per token, which grows with corpus size far faster than a dense index.
- Limits
- LateOn: English, 32-token queries and 300-token documents by default. mLateOn: up to 8,192 tokens, nine training languages. No distillation or prompts were used, which LightOn says leaves room for stronger versions.
What’s good
- 57.22 on BEIR at 149M parameters; 60.36 on decontaminated BEIR
- Apache 2.0 weights and released training data
- mLateOn reads 8,192 tokens in nine languages
- Works as first-stage retriever or reranker in PyLate
What’s not
- Multi-vector indexes are far larger than dense ones
- LateOn is English-only and defaults to 300-token documents
- No hosted API listed for these models
- Benchmarks are LightOn's; no BRIGHT or reasoning results
Best for
Engineers self-hosting retrieval who want a permissively licensed ColBERT model for first-stage search or as a light reranker.
LateOn and mLateOn alternatives
LateOn and mLateOn: common questions
- Is LateOn and mLateOn worth it?
- The strongest open ColBERT models you can use commercially without asking anyone, provided you accept a bigger index in exchange for sharper ranking. We rate it 4.2 out of 5.
- How much does LateOn and mLateOn cost?
- Free weights under Apache 2.0, including commercial use; hosting and index storage are your own cost.
- What are the downsides of LateOn and mLateOn?
- Multi-vector indexes are far larger than dense ones; lateOn is English-only and defaults to 300-token documents; no hosted API listed for these models; benchmarks are LightOn's; no BRIGHT or reasoning results.
- Who is LateOn and mLateOn best for?
- Engineers self-hosting retrieval who want a permissively licensed ColBERT model for first-stage search or as a light reranker.
- How much does LateOn and mLateOn cost for a team or in production?
- No licence or API fee. The running cost is GPU time for encoding and storage for one 128-dimension vector per token, which grows with corpus size far faster than a dense index.
- Is LateOn and mLateOn safe for company data?
- Self-hosted only, so documents stay on your own hardware. Training data for the experiments is public on Hugging Face; the final models used a proprietary Apache-compatible dataset.
- What are the best LateOn and mLateOn alternatives?
- The closest alternatives we have reviewed are LFM2.5-ColBERT-350M, Nemotron ColEmbed V2, Jina Embeddings v5 and Qwen3-Embedding and Qwen3-Reranker.
Answers checked Nov 20, 2026.