OpenVLM CCBench - Sketch Reasoning: leaderboard
CCBench Sketch Reasoning accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): multiple-choice questions on Chinese culture about reasoning about sketches. The category holds 45 of the 510 questions and is scored with CircularEval (correct under every rotation of the options). Accuracy in percent, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 228 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-78B | 95.6 | #361 |
| 2 | InternVL3-8B | 93.3 | #624 |
| 3 | InternVL3-38B | 93.3 | #413 |
| 4 | InternVL3-14B | 93.3 | #506 |
| 5 | InternVL2-8B | 91.1 | #813 |
| 6 | Qwen 2 VL 7B | 86.7 | #743 |
| 7 | Qwen 2 VL 72B | 86.7 | #450 |
| 8 | GPT-4.1 Mini | 84.4 | #363 |
| 9 | GPT-4.1 (2025-04-14) | 84.4 | #257 |
| 10 | MiniCPM-V-2.6 | 84.4 | #827 |
| 11 | Claude 3.7 Sonnet | 80 | #259 |
| 12 | Grok 2 (1212) | 80 | #502 |
| 13 | Moonshot v1 8K | 80 | #622 |
| 14 | Qwen-VL-Max | 80 | #475 |
| 15 | Llama 3.2 90B Vision Instruct | 77.8 | #621 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-ccbench-sketch-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-05.