MinerU 2.5 1.2B

MinerU2.5 is a 1.2B-parameter vision-language model (VLM) for document parsing that achieves state-of-the-art recognition accuracy with high computational efficiency. It employs a coarse-to-fine, two-stage parsing strategy: first performing efficient global layout analysis on downsampled images to identify structural elements, then conducting fine-grained content recognition on native-resolution crops for dense text, complex formulas, and tables. This decoupled approach circumvents the computational overhead of processing high-resolution inputs while preserving fine-grained detail.

Key Strengths

• Comprehensive layout analysis: Preserves non-body elements such as headers, footers, and page numbers, and uses a refined labeling schema for clearer representation of lists, references, and code blocks.
• Formula parsing: Handles complex, lengthy mathematical formulae and accurately recognizes mixed-language (Chinese-English) equations.
• Table parsing robustness: Effectively processes rotated tables, borderless tables, and tables with partial borders.

Benchmark Results

On the olmOCR-bench benchmark, MinerU2.5 achieves an overall score of 75.2, with 76.6 on Arxiv Math, 54.6 on Old Scans Math, and 84.9 on Table Tests. The model is supported by a large-scale, diverse data engine for both pretraining and fine-tuning, enabling robust performance across document types while maintaining low computational overhead.