Benchmark
Parsing quality, measured and verifiable.
Closed parsing APIs ask you to trust their accuracy – they won't tell you which model runs under the hood, let alone how it actually performs. So "best-in-class quality" is a claim you have to verify yourself with tedious evals.
ParseRouter takes the opposite position. The model we run is named, and its quality is a matter of public record. The numbers below come from OmniDocBench, a peer-reviewed parsing benchmark (CVPR 2025) and one of the most extensive available: 1,651 hand-annotated PDF pages across ten document types, from academic papers to financial reports to handwritten notes, scored on text, tables, and formulas against verified ground truth.
MinerU, the model we run today, currently leads it. More hosted models will follow soon.
| Model | Method | Overall ↑ |
|---|---|---|
| MinerU2.5-ProParseRouter | Specialized VLM | 95.75 |
| GLM-OCR | Specialized VLM | 95.22 |
| PaddleOCR-VL-1.5 | Specialized VLM | 94.93 |
| PaddleOCR-VL | Specialized VLM | 94.18 |
| Qianfan-OCR | Specialized VLM | 93.90 |
| Youtu-Parsing | Specialized VLM | 93.74 |
| Ovis2.6-30B-A3B | General VLM | 93.70 |
| Logics-Parsing-v2 | Specialized VLM | 93.33 |
| FireRed-OCR | Specialized VLM | 93.26 |
| MinerU-2.5 | Specialized VLM | 93.04 |
| Gemini 3 Pro | General VLM | 92.91 |
| Gemini 3 Flash | General VLM | 92.62 |
Overall score on OmniDocBench (v1.6, full set). Higher is better. Published by OpenDataLab, the team behind MinerU; dataset and methodology linked so you can verify every number.