mteb-112-languages

IN premise — summaries/2026/08/24/muennighoff-2022-mteb-s5-conclusion.md

Created 2026-08-25T02:58:21+00:00

MTEB covers 112 languages across its multilingual datasets

Summary

The MTEB benchmark pulls its multilingual test data from 112 different languages, so any performance comparison made with it spans an extremely wide range of human languages rather than just English or a few major ones. This gives the system a broad empirical foundation when judging how well embedding models actually work across diverse linguistic contexts in the real world.