CMMLU: leaderboard
11,528 four-option multiple-choice questions across 67 subjects in Chinese, from MBZUAI, LibrAI and Shanghai Jiao Tong University (2023); accuracy, with a random baseline of 25%.
Metric: 5-shot Avg Accuracy (%). Source: github.com. Status: saturated. 42 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 2 72B | 89.65 |
| 2 | Qwen 1.5 110B | 88.32 |
| 3 | Qwen 2.5 72B | 85.67 |
| 4 | Qwen 1.5 72B | 83.54 |
| 5 | Qwen 1.5 32B | 82.25 |
| 6 | Qwen 1.5 7B | 72.5 |
| 7 | XuanYuan-70B | 71.1 |
| 8 | GPT-4 | 70.95 |
| 9 | Llama 3.1 70B Instruct | 64.38 |
| 10 | Qwen-7B | 58.66 |
| 11 | ChatGLM2 6B | 48.8 |
| 12 | falcon-40B | 41.45 |
| 13 | LLaMA-65B | 39.8 |
| 14 | Llama 3 70B Instruct | 36.85 |
Interactive version: theaggregate.ai/benchmark?slug=cmmlu · How It Works · Data refreshed daily, snapshot 2026-09-05.