mLateOn is a multilingual retrieval model that combines a multilingual mmBERT encoder with ColBERT-style late-interaction token matching. Evaluated on the new HAKARI-Bench, it ranked first among eleven late-interaction systems tested despite having only 115 million active parameters, roughly one-sixtieth the size of comparable 8-billion-parameter dense embedding models. It also scored competitively on the MNanoBEIR benchmark, landing between two much larger dense embedding systems while holding up well on long-document and short-query retrieval.
