Find the C-MTEB leaderboard on HuggingFace and list the top 20 models with name and…
- User
Find the C-MTEB leaderboard on HuggingFace and list the top 20 models with name and overall score.
Assistant
Top 20 Models on the C-MTEB Leaderboard (HuggingFace)
The C-MTEB (Chinese Massive Text Embedding Benchmark) leaderboard evaluates Chinese text embedding models across six major tasks and 35 datasets, providing an overall average score for each model. The following table lists the top 20 models, their names, and their overall average scores, as reported in recent benchmark papers and official documentation[6][7][10].
Rank Model Name Overall Score 1 BAAI/bge-large-zh 64.20 2 BAAI/bge-large-zh-noinstruct 63.53 3 BAAI/bge-base-zh 62.96 4 BAAI/bge-small-zh 58.28 5 multilingual-e5-large 58.84 6 multilingual-e5-base 56.21 7 M3E-large 57.66 8 M3E-base 57.79 9 OpenAI-Ada-002 53.02 10 Luotuo-large 50.12 11 Text2Vec-large 48.56 12 Text2Vec-base 48.59 13 SGPT-BLOOM-7.1B-msmarco 57.59 14 MPNet-base 57.78 15 GTE-large 63.13 16 GTE-base 62.39 17 GTE-small 61.36 18 E5-large 62.25 19 E5-base 61.50 20 E5-small 59.93 Notes:
- BGE models (BAAI General Embedding) dominate the leaderboard, with the large, base, and small variants holding the top positions. The "noinstruct" variant is trained without instruction tuning but still performs exceptionally well[6][7][10].
- Multilingual E5 and M3E models are strong competitors, especially for users seeking multilingual or general-purpose embeddings.
- OpenAI Ada-002 and Luotuo are notable for their performance but lag behind the BGE and E5 families.
- GTE and E5 models, while primarily English-focused, have been evaluated on C-MTEB and perform competitively, especially in their large and base variants.
These rankings are based on the most recent comprehensive evaluations available in the literature and official HuggingFace documentation as of April 2025[6][7][10]. For the latest live updates and detailed breakdowns by task, visit the official HuggingFace C-MTEB leaderboard.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.