MinerU 2.5 Pro

MinerU2.5-Pro-2605 is a vision-language model for document parsing that converts PDF pages to structured Markdown, combining layout detection, table and formula recognition, and image/chart analysis in a single 1.2B-parameter architecture.

Key Strengths

The 2605 version improves upon the previous 2604 release by substantially reducing category misclassification in layout detection—especially for the image_block category—and by enhancing recognition of charts, flowcharts, and seals through a large-scale training dataset. These refinements focus on real-world usability while maintaining competitive benchmark performance.

Benchmark Results

On OmniDocBench v1.6full, MinerU2.5-Pro-2605 achieves an overall score of 95.72. The table below compares it to the prior version:

Model Version	Overall↑	Text↓	Formula↑	Table↑	Table↑	Read Order↓
MinerU2.5-Pro-2605	95.72	0.036	97.15	93.62	96.01	0.123
MinerU2.5-Pro-2604	95.69	0.036	97.29	93.42	95.92	0.120

Additional modality-specific results include a formula CDM of 97.15 and a text Edit Distance of 0.036, both at state-of-the-art levels.

Technical Approach

Performance gains come from data engineering rather than architectural changes. The training corpus was expanded from under 10M to 65.5M pages using Diversity-and-Difficulty-Aware Sampling, and annotation quality was improved via Cross-Model Consistency Verification and a Judge-and-Refine pipeline. A three-stage training strategy (large-scale pre-training, hard-sample fine-tuning, GRPO alignment) maximizes data utility. When served with a vllm-async-engine on an A100, the model achieves an inference speed of 2.12 fps.