LightOn's 115M-Parameter mLateOn Tops Multilingual Retrieval Benchmarks

LightOn released mLateOn, a multilingual ColBERT-style retrieval model with only 115M active parameters that ranks first among 11 late-interaction models on HAKARI-Bench, beating the runner-up by 8.3 points with a quarter of the parameters. It also achieves SOTA on unseen languages, rivaling much larger dense models and rerankers.

2026-08-20 ~ 2026-08-21 · 4 related posts

• mLateOn tops multilingual ColBERT retrieval with 115M active params, rivals 8B dense models — IgorCarron · 2026-08-20
• mLateOn wins HAKARI-Bench by 8.33 points with a quarter of runner-up's active params — IgorCarron · 2026-08-20
• LightOn releases mLateOn: SOTA multilingual retrieval with just 115M parameters — antoine_chaffin · 2026-08-20
• 115M parameter model mLateOn achieves multilingual retrieval SOTA — lateinteraction · 2026-08-21