mLateOn Hits SOTA on MTEB Korean Retrieval Without Any Korean Training Data

IgorCarron · x · 2026-08-07

antoinechaffin shared test results of the mLateOn model on the MTEB Korean retrieval task. Despite not being trained on any Korean data, it achieves state-of-the-art performance, outperforming much larger models and getting close to a specialized 8B Korean model. The author notes that this impressive out-of-domain performance is exactly what late interaction mechanisms provide.

Original post →

More from Models

Models channel →