---

license: apache-2.0

base_model:

- opendatalab/MinerU2.5-Pro-2605-1.2B

base_model_relation: quantized

pipeline_tag: image-text-to-text

library_name: gguf

language:

- zh

- en

tags:

- ocr

- document-parsing

- gguf

- llama.cpp

- 42model

---

<!-- README for the Hugging Face repo 42ailab/MinerU2.5-Pro-2605-1.2B-GGUF. Editing this file changes the repo landing page. -->

MinerU2.5-Pro · Fully Local Document Parsing and OCR





English | 简体中文 · Also on ModelScope · 42ailab.com

Give it an image, a scan or a PDF screenshot and this model recognizes the text, tables, formulas and layout structure inside, restoring it to editable, copyable content — entirely on your own computer. No cloud, free, private.

> The model was built by OpenDataLab and released under Apache-2.0 (MinerU2.5-Pro, a 1.2-billion-parameter multimodal document-parsing model). This repo is not a new model — we converted that open model into a quantized build (GGUF) that runs offline on your computer and verified its quality. The document-parsing capability belongs to OpenDataLab; what we did was make it run locally.

1. The problem it solves

Turning paper, scanned or photographed documents back into editable content is the first step of almost any document workflow, and the hard part was never just recognizing characters:

• Tables and formulas: the tables and mathematical formulas in financial reports, papers and contracts are routinely torn apart by ordinary OCR;
• Layout and reading order: multi-column and mixed text-image documents are only useful once restored into the order a human would read them in;
• Whether to go to the cloud: contracts, medical records and internal material carry privacy concerns that make uploading to someone else's server uncomfortable.

This model is built for document parsing: it not only recognizes text but restores tables, formulas and layout structure into structured content, and it does so entirely locally — offline, free, with your originals never leaving your device.

2. How it works

The capability comes from OpenDataLab's MinerU2.5-Pro — a 1.2-billion-parameter multimodal document-parsing model. It follows a "understand the whole page layout first, then read each block closely" approach, parsing a page into structured content (body text, tables, formulas, reading order). Its distinguishing trait is achieving its lead through data engineering rather than parameter count — at just 1.2B, it reaches the top tier of public document-parsing benchmarks (see below).

What we did:

• Converted the open weights into two quantized files — a text decoder of about 0.38 GB (Q4_K_M quantization) plus a vision encoder of about 1.27 GB that does the "looking" — so it runs offline on consumer hardware across macOS / Windows / Linux;
• Verified that quantization introduced no visible quality loss (see below) and included it in the 42model engine's curated OCR selection;
• Kept the same Apache-2.0 license as upstream.

3. How well it performs

For the full public-benchmark evaluation of the model itself, refer to the OpenDataLab technical report (arXiv). The table below quotes the key results for the 2605 release on OmniDocBench v1.6, the authoritative document-parsing benchmark, from that official report (quoted only, not measured by us):

| Benchmark (OpenDataLab official report) | Result |

|---|---|

| OmniDocBench v1.6 overall score (higher is better) | 95.72 (document-parsing SOTA) |

| Table restoration TEDS (higher is better) | 93.62 |

| Formula recognition CDM (higher is better) | 97.15 |

The official report states that on this benchmark it surpasses top specialized OCR models (GLM-OCR, PaddleOCR-VL) and even substantially larger general-purpose multimodal models.

Our own verification (quantization fidelity): we sampled 100 random document images from OmniDocBench v1.6 and compared the full-precision version against this repo's quantized version using the same image and instruction (deterministic output):

| What we verified | Result |

|---|---|

| Character-for-character identical to the full-precision version | 64 / 100 |

| Output length difference < 10% | 90 / 100 |

| Systematic quality degradation | None |

The few differences are character-level fluctuations in full-width/half-width forms and spacing (which the full-precision version also produces between two of its own runs), not artifacts of quantization — in other words, compressed to roughly 40% of the size, with no visible loss in document-parsing quality.

4. Limitations and what's next

• One page at a time: the model's context is about 8192 tokens per call, which fits a single page or image; process multi-page PDFs page by page (the 42model engine feeds them in automatically).
• Built for document parsing, not casual text grabbing: it excels at structured restoration of tables, formulas and complex layouts; if you just want a few words out of a photo, a lighter general-purpose OCR will feel smoother.
• The vision encoder is required: the text decoder and the vision encoder are both mandatory (they are fetched together on download).
• Next: keep verifying against real documents; evaluate more tiers and broader language coverage.

5. How to download and use it

This model is packaged for 42model, which is the recommended way to get it:

Desktop app: open Model Library → OCR → More models, download MinerU2.5-Pro, then click Start to parse documents.

Files and license

| File | Role |

|---|---|

| MinerU2.5-Pro-2605-1.2B-Q4_K_M.gguf | Text decoder (Q4_K_M quantization, ~0.38 GB) |

| mmproj-MinerU2.5-Pro-2605-1.2B-f16.gguf | Vision encoder (the "eyes" that read the image, F16 precision retained, ~1.27 GB) |

Both files are required: the decoder produces text and structure, the vision encoder does the seeing. Per-file sha256 values are listed on the Files page and can be verified independently.

License: the model itself is MinerU2.5-Pro, © OpenDataLab, Apache-2.0 (official sources: Hugging Face · ModelScope · technical report). This repo is a GGUF build converted from it and is likewise governed by Apache-2.0. By using it you agree to the upstream license terms.

Citation

For the model itself, please cite upstream OpenDataLab (the MinerU2.5-Pro technical report):

@misc{wang2026mineru25propushinglimitsdatacentric,
      title={MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing at Scale},
      author={Bin, Wang and Tianyao, He and Linke, Ouyang and Fan, Wu and Zhiyuan, Zhao and Tao, Chu and Yuan, Qu and Zhenjiang, Jin and Weijun, Zeng and Ziyang, Miao and Bangrui, Xu and Junbo, Niu and others},
      year={2026},
      eprint={2604.04771},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.04771}
}

If this repo's GGUF build was useful to you, you may additionally cite:

@misc{yang2026mineru25progguf,
  title        = {MinerU2.5-Pro-2605-1.2B-GGUF: A Local Quantized Build of OpenDataLab's MinerU2.5-Pro},
  author       = {Yang, Zhiping},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/42ailab/MinerU2.5-Pro-2605-1.2B-GGUF}},
  organization = {42ailab},
  note         = {GGUF quantization and local deployment packaging; the model itself is OpenDataLab MinerU2.5-Pro (Apache-2.0). Contact: contact@42ailab.com}
}

Contact us: contact@42ailab.com

About us

42ailab — an AI research lab exploring the boundaries of intelligence. Grounded in cognitive science, we work toward a deep integration of AI and human intelligence — to truly understand and augment intelligence, carbon-based and silicon-based alike.

42model — a high-performance local inference engine from 42ailab that runs translation, transcription, recognition, chat and coding on your own machine, free and private; with optional cloud compute for fine-tuning your own models and bringing them back to run locally.