OCRBench — Band (702-885) matches native ~0-900pt raw scale, not percentage; kept apart.
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | Qwen2.5 VL 72B | Alibaba | 885 | 1 | 2026-06-05 |
| 2 | Kimi VL A3B Thinking | Moonshot | 869 | 2 | 2026-06-12 |
| 3 | Kimi VL A3B Instruct | Moonshot | 867 | 2 | 2026-06-12 |
| 4 | GPT-4o | OpenAI | 815 | 2 | 2026-06-05 |
| 5 | GPT-4o mini | OpenAI | 785 | 1 | 2026-06-05 |
| 6 | Gemma 3 27B | 753 | 1 | 2026-06-05 | |
| 7 | Gemma 3 12B | 702 | 2 | 2026-06-05 |
Best tracked score per model (default configuration; source-attributed and verification-tiered). Open a model for its full benchmark surface, provenance and pricing.