MTEB Leaderboard (Massive Text Embedding Benchmark)

What MTEB is

MTEB, the Massive Text Embedding Benchmark, is an open-source toolkit for evaluating embedding models: the models that turn text into vectors for search, retrieval-augmented generation (RAG), clustering and recommendations. Its public leaderboard on Hugging Face is where most people compare embedding models before choosing one.

What it tests

MTEB groups its tasks into families, including:

• Retrieval: finding the right passage for a query, the task that matters most for RAG.
• Reranking: ordering candidate passages by relevance.
• Classification and clustering: grouping texts by topic or label.
• Semantic textual similarity: scoring how close two sentences are in meaning.
• Bitext mining: matching translations across languages.

The project was expanded in 2025 with MMTEB, the Massive Multilingual Text Embedding Benchmark, which added many more languages and tasks. That makes it far more useful if your users write in Urdu, Hindi, Bengali, Arabic or Swahili, not only English.

How to read the leaderboard

Do not sort by the overall average and take the top row. Instead:

1. Choose the benchmark or language set closest to your users.
2. Sort by the retrieval score if you are building RAG or search.
3. Note each model's vector size and maximum input length: larger vectors cost more storage and slower search.
4. Check whether the model is an API or open weights you can self-host, and its licence.
5. Shortlist three, then test them on 30 to 50 of your own real queries.

Remember that switching embedding models later means re-embedding all your data, so this choice deserves more care than its small price suggests.

Run it yourself

The toolkit installs with pip install mteb and can evaluate any model, including one you fine-tuned, on the same tasks.

Free or paid?

The leaderboard and the toolkit are free and open source under the Apache 2.0 licence.

Common questions

Is the top model always the best choice? No. The top rows are often large models with big vectors. A smaller model a few points lower may be cheaper to run, faster to search and just as good on your own queries.

Best for: anyone choosing an embeddings model for search, RAG or deduplication, especially in languages other than English.