mLateOn

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

🎯 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)

MIRACLtgt and MLDRtgt include only the languages overlapping with our target languages; MIRACL and MLDR include all benchmark languages.