Modelpublic

Abiray/Qianfan-OCR-GGUF

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face

1likes387downloads

Model Card

Model Card for Qianfan-OCR-4B-GGUF

This repository contains GGUF format quantizations (Q4KM, Q5KM, Q6K, Q80) of Qianfan-OCR, a 4B-parameter end-to-end document intelligence model developed by the Baidu Qianfan Team. It unifies document parsing, layout analysis, and document understanding within a single vision-language architecture.

Model Details

Model Description

Qianfan-OCR is designed to replace traditional multi-stage OCR pipelines. Instead of chaining separate layout detection and text recognition modules, it performs direct image-to-Markdown conversion. It introduces Layout-as-Thought, an optional thinking phase where the model generates structured layout representations before producing the final output.

—Developed by: Baidu Qianfan Team
—Shared by: Abhiray
—Model type: Multimodal Vision-Language Model (VLM)
—Text Backbone: Qwen3-4B
—Vision Encoder: Qianfan-ViT (AnyResolution up to 4K)
—Language(s) (NLP): Multilingual (192 languages supported)
—License: Apache-2.0

Model Sources

—Repository: Baidu Qianfan-OCR
—Paper: Qianfan-OCR: A Unified End-to-End Model for Document Intelligence

Uses

Direct Use

—Document Parsing: High-fidelity Image-to-Markdown conversion.
—Table/Formula Recognition: Extracting complex tables (merged cells) and LaTeX formulas.
—Key Information Extraction (KIE): Structured data extraction from receipts, invoices, and IDs.
—Visual Question Answering (DocVQA): Reasoning over charts, graphs, and structured documents.

Out-of-Scope Use

—The model is not intended for generating creative fiction or conversational roleplay; it is optimized for high-accuracy document intelligence and visual grounding.

How to Get Started with the Model

1. Requirements

You need both a Quantized Model File (.gguf) and the Vision Projector (qianfan-ocr-mmproj.gguf).

2. Running with Koboldcpp (Recommended)

bash

./koboldcpp \
  --model qianfan-ocr-4b-Q4_K_M.gguf \
  --mmproj qianfan-ocr-mmproj.gguf \
  --usecuda 0 \
  --gpulayers 99 \
  --contextsize 8192 \
  --remotetunnel