--- title: Logics-Parsing-V3 canonical_url: "https://www.modelscope.cn/models/Alibaba-DT/Logics-Parsing-V3" md_url: "https://www.modelscope.cn/models/Alibaba-DT/Logics-Parsing-V3.md" repository: Alibaba-DT/Logics-Parsing-V3 chinese_name: Logics-Parsing-V3 last_updated: 2026-09-22 model_type: - qwen3_5 architectures: - Qwen3_5ForConditionalGeneration parameters: 853.0M tensor_type: - BF16 library_name: - safetensors downloads: 48 stars: 1 --- # Logics-Parsing-V3 > Logics-Parsing-V3 - Alibaba-DT 在 ModelScope 开源的模型。💻 HomePage | 🤗 GitHub | 🤖 Demo - **Repository**: Alibaba-DT/Logics-Parsing-V3 - **Parameters**: 853.0M - **Downloads**: 48 - **Stars**: 1 - **Last updated**: 2026-09-22 Source: https://www.modelscope.cn/models/Alibaba-DT/Logics-Parsing-V3 ---

💻 HomePage | 🤗 GitHub | 🤖 Demo

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](#evaluation) for the full results and document-length analysis.

Overall performance comparison on MPDocBench

Metric-level radar comparison on MPDocBench

## News - **2026/09/22:** 🚀 Released Logics-Parsing-V3. - **2026/03/09:** Released [Logics-Parsing-Omni](https://github.com/alibaba/Logics-Parsing/tree/main/Logics-Parsing-Omni). See the [technical report](https://arxiv.org/pdf/2603.09677) 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.

Model	Document-Level Joint Parsing	Cross-Page Element Merging	Visual–Text Association	Hierarchy Recovery
MinerU2.5-Pro	✗	✓	✗	✗
MinerU2.5-Pro + MinerU-PoPo	✗	✓	✓	✓
Unlimited-OCR	✓	✗	✗	✗
Logics-Parsing-V3	✓	✓	✓	✓

### 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.

Parsing performance by document length on MPDocBench-Long

### 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.

Model	Total Time (min) ↓	Time per Page (s) ↓
OvisOCR2	73.26	1.40
Unlimited-OCR	90.12	1.72
MinerU2.5-Pro	175.10	3.35
MinerU2.5-Pro + MinerU-PoPo	197.49	3.78
Logics-Parsing-V3	72.46	1.39

*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](https://huggingface.co/Logics-MLLM/Logics-Parsing-V3/tree/main) (default) or [ModelScope](https://www.modelscope.cn/models/Alibaba-DT/Logics-Parsing-V3). 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

Multilevel document structure comparison on a Chinese textbook

#### Cross-Page Tables

Additional cross-page table merging example

#### Cross-Page Text Continuity

### Single-Page Parsing #### Academic Posters: Reading Order

Reading order recovery in an academic poster

#### Complex Tables

#### Code Blocks

#### Chemistry Exam Papers

Chemical notation recognition in an exam paper

#### Handwritten Notes