C-Eval — Results

Rank	Model	Score
1	deepseek-v4-1-flash	92.1
2	ling-3-0-flash-base	89.76
3	ling-3-0-tiny-base	80.59
4	qwen3-6-plus	0.933
5	qwen3-5-397b-a17b	0.93
6	kimi-k2-base	0.925
7	qwen3-5-122b-a10b	0.919
8	mimo-v2-5-pro	0.915
9	qwen3-6-27b	0.914
10	qwen3-5-27b	0.905
11	qwen3-5-35b-a3b	0.902
12	qwen3-6-35b-a3b	0.9
13	kimi-k1-5	0.883
14	qwen3-5-9b	0.882
15	deepseek-v3	0.865
16	qwen3-5-4b	0.851
17	qwen2-72b-instruct	0.838
18	qwen2-7b-instruct	0.772
19	qwen3-5-2b	0.732
20	qwen3-5-0-8b	0.505
21	ernie-4-5	0.407

C-Eval

1 phaseActive

Comprehensive Chinese evaluation suite with 13,948 MCQs across 52 disciplines at four difficulty levels: middle school, high school, college, and professional.

Overview

C-Eval

Quick answer: C-Eval (Huang et al., 2023) is a comprehensive Chinese evaluation suite with 13,948 multiple-choice questions across 52 disciplines at four difficulty levels. It assesses advanced Chinese knowledge and reasoning across humanities, science, and engineering.

Benchmark Details

Property	Value
Tasks	13,948 multiple-choice questions
Disciplines	52 (humanities, science, engineering)
Difficulty levels	Middle school, high school, college, professional
Language	Chinese
Metric	Accuracy
Subset	C-Eval Hard (very challenging subjects)

Source: Huang et al. 2023. Last updated 2026-07-24.