OpenVLM ChartQA Test - Human: leaderboard
ChartQA test human split from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): questions about charts written by people, which often need visual and logical reasoning, scored by relaxed accuracy (numeric answers within 5% count). Percent, higher is better; the OpenVLM ChartQA Test board averages the two splits, and an all-zero failed run is dropped.
Metric: Relaxed Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 108 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2 VL 72B | 81.8 | #450 |
| 2 | InternVL3-14B | 80.5 | #506 |
| 3 | InternVL3-8B | 77 | #624 |
| 4 | MiniCPM-V-2.6 | 75 | #827 |
| 5 | Pixtral-12B | 62.2 | #805 |
| 6 | GPT-4.1 Mini | 38.4 | #363 |
| 7 | GPT-4.1 (2025-04-14) | 37.3 | #257 |
| 8 | GPT-4.1 Nano | 34.2 | #730 |
| 9 | Gemini 1.5 Flash | 32.8 | #620 |
| 10 | Gemma 3 12B | 31.9 | #674 |
| 11 | Aquila-VL-2B | 27.2 | #1216 |
| 12 | Gemma 3 27B | 26.6 | #606 |
| 13 | Gemma 3 4B | 24.5 | #1088 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-chartqa-test-human · How It Works · Data refreshed daily, snapshot 2026-11-05.