OmniDocBench
5 models tested · Updated 2026-04-06 · Verified sources only
MinerU2.5-Pro leads at 95.69%
1
Shanghai AI Lab · arxiv/2604.04771 · 2026-04-06
SOTA on OmniDocBench v1.6 via pure data engineering — no architecture changes. Expanded training from <10M to 65.5M samples.
95.69%
2
Zhipu AI · arxiv/2603.10910 · 2026-03-11
0.9B model (0.4B encoder + 0.5B decoder) achieves #1 on OmniDocBench v1.5. Outperforms models 260x larger including GPT-5.2 and Gemini-3 Pro.
94.62%
3
Baidu · arxiv/2601.21957 · 2026-01-29
0.9B compact VLM with NaViT dynamic resolution. Best text and formula recognition scores despite minimal size. Also introduces Real5-OmniDocBench for real-world robustness.
94.5%
4
Baidu · arxiv/2603.13398 · 2026-03-11
End-to-end unified OCR model. Wins 6/8 benchmarks in its class. Strong chart reasoning capabilities.
93.12%
5
FireRed AI · arxiv/2603.01840 · 2026-03-02
3-stage training: multi-task pre-alignment, specialized SFT, and GRPO RL with format constraints. Qwen3-VL backbone.
92.94%