OCRBench v2 (English)
Multimodal · 2024-12-31
The English portion of OCRBench v2's public set: questions over 21 task types grouped into eight capabilities (text recognition, referring, spotting, relation extraction, element parsing, mathematical calculation, visual text understanding and knowledge reasoning). The official English score averages item scores within each capability and then takes the unweighted mean of the eight; the private leaderboard set is a different cohort.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemma 4 12B IT | 67.2 |
| Qwen3.5-4B | 65.7 |
| GPT-5.6 Luna | 64.8 |
| LFM3-VL-2B | 63.5 |
| Gemini 3.5 Flash-Lite | 55.5 |
| Qwen3.5-2B | 55.4 |
| Gemma 4 E2B IT | 51.3 |
| LFM2.5-VL-3B | 47.5 |