Abstract

We introduce MinerU2.5, a 1.2B-parameter document parsing vision-languagemodel that achieves state-of-the-art recognition accuracy while maintainingexceptional computational efficiency. Our approach employs a coarse-to-fine,two-stage parsing strategy that decouples global layout analysis from localcontent recognition. In the first stage, the model performs efficient layoutanalysis on downsampled images to identify structural elements, circumventingthe computational overhead of processing high-resolution inputs. In the secondstage, guided by the global layout, it performs targeted content recognition onnative-resolution crops extracted from the original image, preservingfine-grained details in dense text, complex formulas, and tables. To supportthis strategy, we developed a comprehensive data engine that generates diverse,large-scale training corpora for both pretraining and fine-tuning. Ultimately,MinerU2.5 demonstrates strong document parsing ability, achievingstate-of-the-art performance on multiple benchmarks, surpassing bothgeneral-purpose and domain-specific models across various recognition tasks,while maintaining significantly lower computational overhead.