OpenVLM OCRBench - Text Recognition: leaderboard
OCRBench Text Recognition from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): reading words from regular, irregular, artistic, handwritten, digit-string and non-semantic text images. This part holds 300 of OCRBench's 1,000 questions; the board shows correct answers as a percentage of those 300, higher is better.
Metric: Correct Answers (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 285 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.5 | 94 | #278 |
| 2 | InternVL3-38B | 93.33 | #413 |
| 3 | InternVL3-8B | 92.33 | #624 |
| 4 | InternVL3-78B | 92.33 | #361 |
| 5 | GPT-4.1 (2025-04-14) | 92 | #257 |
| 6 | GPT-4o ChatGPT | 91.33 | |
| 7 | GPT-5 | 90.67 | #105 |
| 8 | InternVL3-14B | 90.67 | #506 |
| 9 | GPT-5 Mini (2025-08-07) | 90.67 | #175 |
| 10 | GPT-4.1 Mini | 90 | #363 |
| 11 | Qwen 2.5 VL 7B | 89.33 | #641 |
| 12 | Claude 3.5 Sonnet (20241022) | 89 | #302 |
| 13 | Qwen 2 VL 72B | 89 | #450 |
| 14 | Moonshot v1 8K | 89 | #622 |
| 15 | Gemini 2.0 Flash | 88.67 | #351 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-ocrbench-text-recognition · How It Works · Data refreshed daily, snapshot 2026-11-05.