MTEB (Massive Text Embedding Benchmark) is a standardised evaluation suite that measures text embedding models across many tasks, including retrieval, classification, clustering, reranking, and semantic similarity, over numerous datasets and languages. It provides a public leaderboard that has become the reference for comparing embedding models. Strong MTEB scores are widely used to select embeddings for semantic search and retrieval-augmented generation.
Content
- By aggregating dozens of tasks and datasets into a single leaderboard, MTEB exposes how embeddings generalise beyond any one objective, discouraging overfitting to a single benchmark. Practitioners consult it to balance retrieval quality, multilinguality, dimensionality, and inference cost when choosing an embedding model.