---

license: apache-2.0

base_model:

- Logics-MLLM/Logics-Parsing-V3

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/Logics-Parsing-V3-GGUF. Editing this file changes the repo landing page. -->

Logics-Parsing-V3 · A 0.8B Document Parser That Reads Layout, Not Just Characters

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




Give it a page — a scanned book, a paper, a photo of a document — and it returns structured text: headings as headings, body as body, tables as tables, formulas as formulas. All of it runs on your own machine: no upload, free, private.

> The model was built by the Alibaba Logics team and released under Apache-2.0 (Logics-Parsing-V3, a Qwen3.5 vision-language architecture). This repository is not a new model — it is a quantized GGUF build we produced so the model runs offline on your own computer, together with our quality checks. The capability belongs upstream; what we did was make it run locally.

1. The problem it solves

Getting text out of scans, papers and photographed pages usually breaks in three places: layout (two columns, captions and running heads collapse into one stream), formulas (turned into noise or dropped), and tables (merged cells flattened into a single row).

Handling those has usually meant uploading the file to a cloud service — a step people hesitate over with contracts, internal documents and unpublished drafts.

Logics-Parsing-V3 is only 0.8B parameters; quantized it is under 800 MB and runs on an ordinary laptop, so the file never leaves your machine.

2. How it works

The capability comes from the Alibaba Logics team's Logics-Parsing-V3. Its design points:

• State-recurrent parsing across pages. Instead of fitting a whole document into one context window, it carries a compact structural state forward from page to page. Quality therefore degrades more slowly as documents get longer.
• Structure and content in one pass. Heading hierarchy, cross-page merging and figure–text association are produced by the model itself rather than by a chain of rules bolted on afterwards.

What we did: converted the upstream weights into one quantized file (774 MB) plus a vision encoder (195 MB), verified that quantization did not degrade quality, and made it run offline across macOS / Windows / Linux — under the same Apache-2.0 license as upstream.

3. How well it works

For the model's full benchmark results on MPDocBench and OmniDocBench v1.6, please refer to the upstream repository's benchmark figures (GitHub). Below are our own checks, plus the one table upstream publishes as numbers rather than as an image.

Our measurements (macOS · Apple Silicon · Q8_0 + f16 vision encoder):

| What we verified | Result |

|---|---|

| Quantization fidelity (Q8_0 vs Q4_K_M, same page, character by character) | Identical. We therefore ship Q8_0 — for a 0.8B model, saving 269 MB is not worth any precision risk |

| Chinese book page · easily-confused word | Read 你来代入 correctly (two HunyuanOCR builds in our own library read it as 你来代替 on the same page) |

| Chinese book page · mathematical notation | Read P(E｜¬H) correctly and rendered it as \neg (previously only the 7B olmOCR-2 managed this in our library) |

| English paper page | LaTeX superscripts in the author block, plus project and dataset URLs, all preserved verbatim |

⚠️ Evidence strength, stated honestly: the "confused word" and "notation" rows each rest on a single page. They hold for that page and should not be generalized into a capability profile.

Upstream's officially reported efficiency figures (MPDocBench, 420 documents / 3,135 pages, H100, batch size 1, source):

| Model | Total time (min) ↓ | Per page (s) ↓ |

|---|---:|---:|

| Logics-Parsing-V3 | 72.46 | 1.39 |

| OvisOCR2 | 73.26 | 1.40 |

| Unlimited-OCR | 90.12 | 1.72 |

| MinerU2.5-Pro | 175.10 | 3.35 |

4. Limitations and what comes next

• Only single-page mode runs locally today. Upstream also defines a long-document mode that carries structural state across pages; wiring that up requires feeding multiple pages and maintaining state, which our local engine does not do yet. Single-page mode already covers "photograph a page, read a page".
• Figures and charts return a location, not content. The model emits layout boxes for illustrations; with no local image store to put them in, those blocks are skipped. Text is unaffected.
• Heading levels and cross-page continuation markers belong to long-document mode and do not appear in single-page output.

5. How to download and use

This build is packaged for 42model; getting it through 42model is recommended:

Option 1 · Desktop app: open Model Library → OCR, download Logics-Parsing-V3, then click "Start".

Option 2 · Command line (CLI):

42model download logics-parsing-v3

Files and license

| File | Role |

|---|---|

| Logics-Parsing-V3-Q8_0.gguf | Text decoder (Qwen3.5 family) |

| mmproj-Logics-Parsing-V3-f16.gguf | Vision encoder (kept at F16) |

Per-file sha256 values are on the repository's file page and can be verified independently.

License: the model itself is Logics-Parsing-V3, © Alibaba Logics team, Apache-2.0 (official sources: GitHub · Hugging Face · ModelScope). This repository is a GGUF deployment build of it and follows the same Apache-2.0; using it means accepting the upstream license terms.

Citation

For the model itself, please cite the Alibaba Logics team. The Logics-Parsing-V3 technical report is marked "under review" in the upstream repository, so no public citation entry exists yet — please refer to the GitHub repository. The technical report for the sibling Logics-Parsing-Omni is at arXiv:2603.09677.

If this repository's GGUF build is useful to you, you may additionally cite:

@misc{yang2026logicsparsingv3gguf,
  title        = {Logics-Parsing-V3-GGUF: A Local Quantized Build of Alibaba's Logics-Parsing-V3},
  author       = {Yang, Zhiping},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/42ailab/Logics-Parsing-V3-GGUF}},
  organization = {42ailab},
  note         = {GGUF quantization and local-deployment packaging; the model itself is Alibaba's Logics-Parsing-V3 (Apache-2.0). Contact: contact@42ailab.com}
}

Contact us: contact@42ailab.com

About us

42ailab — an AI lab exploring the boundaries of intelligence. Grounded in cognitive science, we work toward a deeper integration of AI and human intelligence: understanding and augmenting intelligence itself — carbon-based and silicon-based alike.

42model — a high-performance local inference engine for AI models, from 42ailab. It lets translation, transcription, recognition, chat and coding run on your own machine, free and private; you can also fine-tune your own model with cloud compute and bring it back to run locally.