MTEB Leaderboard

Overview

MTEB (Massive Text Embedding Benchmark) is a massive benchmark for measuring the performance of text embedding models on diverse embedding tasks. MTEB spans 8 embedding tasks covering a total of 58 datasets and 112 languages.

Expansion: MMTEB

MMTEB (Massive Multilingual Text Embedding Benchmark) expands MTEB to over 500 quality-controlled evaluation tasks across 1,000+ languages, making it the most comprehensive multilingual embedding benchmark.

Current Top Models (2025-2026)

gte-Qwen3-8B

• Largest of a family of new embedding models built on top of Qwen3
• Outperforms previous generation Qwen embedding models
• Ranks high on both multi-lingual and English-only MTEB leaderboards

NV-Embed-v2 (NVIDIA)

• Released October 2025
• Fine-tuned from Llama-3.1-8B
• Particularly powerful at understanding multilingual text
• Previous version (NV-Embed) achieved score of 69.32 on MTEB (56 embedding tasks)

Benchmark Tasks

MTEB covers 8 main embedding tasks:

• Classification
• Clustering
• Pair classification
• Reranking
• Retrieval
• Semantic Textual Similarity (STS)
• Summarization
• Bitext mining

How to Access

Official MTEB leaderboard available at: https://huggingface.co/spaces/mteb/leaderboard

Users can select different benchmarks:

• Multilingual leaderboard
• English-only leaderboard
• Domain-specific benchmarks

Research Impact

MTEB has become the de facto standard for:

• Comparing embedding model performance
• Selecting appropriate models for specific tasks
• Tracking progress in text embedding research
• Validating new embedding approaches