New: connect Claude & other AIs to GenAIList over MCP — research the catalog and contribute to the shared knowledge base. Learn how →
C-Eval benchmark
AI model leaderboard for the C-Eval benchmark. Compare how large language models score on C-Eval, see the full ranking, and understand what this AI benchmark measures. Qwen3-235B-A22B-Thinking (Jul 2025) currently leads with 93.13%. Multi-discipline Chinese exam benchmark across 52 subjects.
Leaderboard
| # | Model | Organization | Score | Variant | Source |
|---|---|---|---|---|---|
| #1 | Qwen3-235B-A22B-Thinking (Jul 2025) | Qwen | 93.13% | thinking-cited-ring-1t | official ↗ |
| #2 | dots.llm1 | Rednote | 92.2% | EM | official ↗ |
| #3 | Ling-1T | Ant Group | 92.19% | official ↗ | |
| #4 | DeepSeek-V3.1-Terminus | DeepSeek | 91.76% | non-thinking-cited-ling-1t | official ↗ |
| #5 | Ring-1T | Ant Group | 91.53% | thinking | official ↗ |
| #6 | DeepSeek-V3.1-Terminus | DeepSeek | 91.22% | thinking-cited-ring-1t | official ↗ |
| #7 | Kimi K2 | Moonshot | 91.12% | instruct-0905-cited-ling-1t | official ↗ |
| #8 | Pangu Pro MoE | Huawei | 91.1% | instruct-em | official ↗ |
| #9 | gemini-2.5-pro | Google DeepMind | 90.14% | cited-ring-1t | official ↗ |
| #10 | Qwen3-235B-A22B | Qwen | 89.6% | thinking | official ↗ |
| #11 | gemini-2.5-pro | Google DeepMind | 88.77% | lowthink-cited-ling-1t | official ↗ |
| #12 | GPT-5 | OpenAI | 88.27% | thinking-high-cited-ring-1t | official ↗ |
| #13 | Qwen2-57B-A14B | Qwen | 87.7% | base-from-hf-readme | official ↗ |
| #14 | Qwen3-32B | Qwen | 87.3% | thinking | official ↗ |
| #15 | Telechat2-115B | China Telecom | 86.9% | from-hf-readme | official ↗ |
| #16 | Qwen3-30B-A3B | Qwen | 86.6% | thinking | official ↗ |
| #17 | DeepSeek-V3 | DeepSeek | 86.3% | official ↗ | |
| #18 | Qwen3-14B | Qwen | 86.2% | thinking | official ↗ |
| #19 | Qwen3-235B-A22B | Qwen | 86.1% | non-thinking | official ↗ |
| #20 | TeleChat2-35B | China Telecom | 85% | from-hf-readme | official ↗ |
| #21 | Qwen2-72B | Qwen | 83.8% | instruct-cited-telechat2 | official ↗ |
| #22 | GPT-5 | OpenAI | 83.59% | main-cited-ling-1t | official ↗ |
| #23 | Qwen1.5-32B | Qwen | 83.5% | cited-qwen2-57b | official ↗ |
| #24 | Qwen3-8B | Qwen | 83.4% | thinking | official ↗ |
| #25 | Qwen3-32B | Qwen | 83.3% | non-thinking | official ↗ |
| #26 | Qwen2-7B | Qwen | 83.2% | base-from-hf-readme | official ↗ |
| #27 | Qwen3-30B-A3B | Qwen | 82.9% | non-thinking | official ↗ |
| #28 | TeleChat2-7B | China Telecom | 82% | from-hf-readme | official ↗ |
| #29 | Qwen3-14B | Qwen | 81% | non-thinking | official ↗ |
| #30 | DeepSeek-V2 (MoE-236B) | DeepSeek | 78% | cited-telechat2 | official ↗ |
| #31 | Qwen3-8B | Qwen | 77.9% | non-thinking | official ↗ |
| #32 | Qwen3-4B | Qwen | 77.5% | thinking | official ↗ |
| #33 | Llama 4 Scout | Meta AI | 77.3% | cited-pangu-pro-moe | official ↗ |
| #34 | TeleChat2-3B | China Telecom | 75% | from-hf-readme | official ↗ |
| #35 | Qwen1.5-7B | Qwen | 74.1% | base-cited-qwen2-7b | official ↗ |
| #36 | Qwen3-4B | Qwen | 72.2% | non-thinking | official ↗ |
| #37 | Qwen2-1.5B | Qwen | 70.6% | base-from-hf-readme | official ↗ |
| #38 | Qwen3-1.7B | Qwen | 68.1% | thinking | official ↗ |
| #39 | Qwen3-1.7B | Qwen | 61% | non-thinking | official ↗ |
| #40 | Qwen2-0.5B | Qwen | 58.2% | base-from-hf-readme | official ↗ |
| #41 | Qwen3-0.6B | Qwen | 50.4% | thinking | official ↗ |
| #42 | Llama 3-8B | Meta AI | 49.5% | base-cited-qwen2-7b | official ↗ |
| #43 | Mistral 7B | Mistral AI | 47.4% | base-cited-qwen2-7b | official ↗ |
| #44 | Gemma 7B | Google DeepMind | 43.6% | base-cited-qwen2-7b | official ↗ |
| #45 | Qwen3-0.6B | Qwen | 42.6% | non-thinking | official ↗ |
Frequently asked questions about C-Eval
What is the C-Eval benchmark?
Multi-discipline Chinese exam benchmark across 52 subjects.
How is the C-Eval benchmark scored?
C-Eval is scored using the accuracy (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.
Which AI model scores highest on C-Eval?
As of the latest reported scores on GenAIList, Qwen3-235B-A22B-Thinking (Jul 2025) achieves the highest result on C-Eval with a score of 93.13%.
Is a higher C-Eval score better?
Yes. On C-Eval a higher score indicates better performance, so models near the top of the leaderboard are the strongest.