Modelpublic

Logics-MLLM/Logics-Parsing-V3

sourceHugging Faceupdated 1d agoView on Hugging Face

6likes24downloads

Model Card

<table align="center" width="100%"> <tr> <td width="10000" align="center"> <a href="imgs/logo.png"><img src="imgs/logo.png" alt="Logics-Parsing-V3" width="740"></a> <br><br> 💻 <a href="https://logics.alibaba-inc.com/parsing/">HomePage</a>   |   🤗 <a href="https://github.com/alibaba/Logics-Parsing/tree/main?tab=readme-ov-file">GitHub</a>   |   🤖 <a href="https://www.modelscope.cn/studios/Alibaba-DT/Logics-Parsing">Demo</a> </td> </tr> </table>

Logics-Parsing-V3 is a 0.8B vision-language model that extends Logics-Parsing-v2 from page-level recognition to structured long-document parsing. It combines global structural understanding, fine-grained local parsing, and scalability to very long documents. By carrying a compact structural state across pages, it preserves document context while focusing on local content, incrementally reconstructing a coherent document tree beyond a single context window. It also retains support for complex layouts and specialized content, including scientific formulas and chemical notation.

Benchmark Results

The charts below summarize overall performance and the metric profile on MPDocBench. See Evaluation for the full results and document-length analysis.

News

—2026/09/22: 🚀 Released Logics-Parsing-V3.
—2026/03/09: Released Logics-Parsing-Omni. See the technical report for details.
—2026/02/13: Released Logics-Parsing-v2.
—2025/09/25: Released Logics-Parsing.

Key Features

1. State-Recurrent Document Modeling

Propagates an evolving structural state across pages, preserving document context in compact memory rather than requiring the full document in a single context window. This recurrent formulation enables incremental structure recovery as document length grows.

2. Hierarchical Document Parsing

Unifies global structural reasoning with fine-grained local recognition to recover a coherent document hierarchy. Reconstructs heading levels, merges cross-page elements, and links visual content with related text—all within a single end-to-end model.

<table align="center"> <thead> <tr><th>Model</th><th>Document-Level Joint Parsing</th><th>Cross-Page Element Merging</th><th>Visual–Text Association</th><th>Hierarchy Recovery</th></tr> </thead> <tbody> <tr><td>MinerU2.5-Pro</td><td align="center">✗</td><td align="center">✓</td><td align="center">✗</td><td align="center">✗</td></tr> <tr><td>MinerU2.5-Pro + MinerU-PoPo</td><td align="center">✗</td><td align="center">✓</td><td align="center">✓</td><td align="center">✓</td></tr> <tr><td>Unlimited-OCR</td><td align="center">✓</td><td align="center">✗</td><td align="center">✗</td><td align="center">✗</td></tr> <tr><td><strong>Logics-Parsing-V3</strong></td><td align="center"><strong>✓</strong></td><td align="center"><strong>✓</strong></td><td align="center"><strong>✓</strong></td><td align="center"><strong>✓</strong></td></tr> </tbody> </table>

3. Leading Performance

With just 0.8B parameters, Logics-Parsing-V3 combines leading multi-page parsing performance on MPDocBench with highly competitive single-page results on OmniDocBench v1.6, while demonstrating strong robustness to increasing document length on MPDocBench-Long.

Evaluation

Multi-Page Parsing: MPDocBench

On MPDocBench (MPDocBench-Parse), Logics-Parsing-V3 achieves the highest Overall score among the compared models, with leading results on Truncated Text Edit and Heading TEDS. These results highlight its strengths in cross-page text continuity and heading hierarchy recovery.

Single-Page Parsing: OmniDocBench v1.6

In addition to directly parsing multi-page documents, Logics-Parsing-V3 also handles single-page documents effectively, achieving competitive results across text, formulas, tables, and reading order on OmniDocBench v1.6.

Length Scaling: MPDocBench-Long

To examine longer inputs beyond the predominantly short documents in MPDocBench, we construct MPDocBench-Long by concatenating complete source documents with similar types and resolutions. The benchmark spans five length ranges up to 50 pages, with 50 sequences per range.

Under the 1M inference setting, Logics-Parsing-V3 maintains more consistent parsing quality as document length increases, while Unlimited-OCR shows a larger decline.

Inference Efficiency

With only 0.8B parameters, Logics-Parsing-V3 combines efficient local parsing with document-level context modeling. It processes pages sequentially using non-overlapping sliding windows, with each window incorporating the current page content and a compact structural state carried forward from the preceding window. This recurrent design preserves cross-page context without repeatedly processing the full document history.

We benchmark inference efficiency on MPDocBench (420 documents, 3,135 pages) using NVIDIA H100 GPUs, with batch size 1 for all models.

<table align="center"> <thead> <tr><th>Model</th><th>Total Time (min) ↓</th><th>Time per Page (s) ↓</th></tr> </thead> <tbody> <tr><td>OvisOCR2</td><td align="right">73.26</td><td align="right">1.40</td></tr> <tr><td>Unlimited-OCR</td><td align="right">90.12</td><td align="right">1.72</td></tr> <tr><td>MinerU2.5-Pro</td><td align="right">175.10</td><td align="right">3.35</td></tr> <tr><td>MinerU2.5-Pro + MinerU-PoPo</td><td align="right">197.49</td><td align="right">3.78</td></tr> <tr><td>Logics-Parsing-V3</td><td align="right"><strong>72.46</strong></td><td align="right"><strong>1.39</strong></td></tr> </tbody> </table> Note: Logics-Parsing-V3 was evaluated with a window size of 3.

Quick Start

1. Installation

bash

conda create -n logics-parsing-v3 python=3.10 -y
conda activate logics-parsing-v3

python -m pip install --upgrade pip
python -m pip install "vllm==0.22.1" "transformers==5.6.0" Pillow PyMuPDF
python -m pip check

2. Download Model Weights

Download Logics-Parsing-V3 from Hugging Face (default) or ModelScope. Install only the download provider you choose.

bash

# Hugging Face
python -m pip install -U huggingface_hub
python download_model.py --source hf

# ModelScope (alternative)
python -m pip install -U modelscope
python download_model.py --source modelscope

3. Inference

inference_v3.py supports images, PDFs, and directories of page images. Multi-page inputs use --window_size 2 by default; single-page inputs always use 1.

Single-page image

bash

python inference_v3.py \
  --model_path ./weights/Logics-Parsing-V3 \
  --input_path /path/to/page.png \
  --output_dir ./outputs/single_page

Multi-page PDF

bash

python inference_v3.py \
  --model_path ./weights/Logics-Parsing-V3 \
  --input_path /path/to/document.pdf \
  --output_dir ./outputs/document

Outputs in --output_dir include parsed content (markdown.md), a multi-page document outline (hierarchy.md), raw labels (raw_output.dsl), and parsing results with metadata (result.json).

Showcases

Long-Document Parsing

Hierarchical Structure

Cross-Page Tables

Cross-Page Text Continuity

Single-Page Parsing

Academic Posters: Reading Order

Complex Tables

Code Blocks

Chemistry Exam Papers

Handwritten Notes