SandLogicTechnologies/DeepSeek-OCR-2-GGUF

DeepSeek-OCR-2

DeepSeek-OCR-2 is a vision-language OCR model developed by DeepSeek, designed for high-quality document parsing, optical character recognition, and structured information extraction from complex visual documents. This repository contains GGUF quantized variants of the model optimized for efficient local inference using llama.cpp.

The model is built to accurately recognize and interpret textual information from a wide variety of document types, including scanned pages, invoices, receipts, forms, tables, books, reports, and handwritten or digitally generated content. Beyond text recognition, it emphasizes document parsing and contextual understanding to enable reliable extraction of structured information from visually complex documents.

The quantized formats significantly reduce memory requirements while maintaining strong OCR accuracy and document comprehension, making the model suitable for both local deployment and enterprise-scale document-processing pipelines.

---

Model Overview

• Model Name: DeepSeek-OCR-2
• Base Model: deepseek-ai/DeepSeek-OCR-2
• Architecture: Vision-Language OCR Model
• Parameter Count: Approximately 3 Billion Parameters
• Modalities: Text, Image
• Primary Languages: Multilingual
• Developer: DeepSeek AI
• License: MIT

---

Quantization Formats

This repository provides various GGUF quantized versions of the DeepSeek-OCR-2 model optimized for efficient local inference using llama.cpp.

IQ3_M

• Size reduction of approx 75.32% (1.35 GB) compared to 16-bit (5.47 GB)
• Aggressive 3-bit quantization optimized for memory-efficient OCR and document parsing workloads
• Suitable for lightweight document digitization systems and edge deployments
• Enables efficient extraction of textual information from diverse document formats on consumer hardware
• Recognition quality for dense layouts, handwritten text, and visually complex documents may reduce compared to higher-precision variants

IQ4_NL

• Size reduction of approx 70.93% (1.59 GB) compared to 16-bit (5.47 GB)
• Advanced 4-bit non-linear quantization designed to better preserve OCR fidelity and document parsing quality
• Better suited for structured document analysis, information extraction, and layout-aware processing workflows
• Designed to reduce quantization loss while maintaining consistent document interpretation
• May require slightly increased computational overhead during inference

IQ4_XS

• Size reduction of approx 72.03% (1.53 GB) compared to 16-bit (5.47 GB)
• Balanced 4-bit quantization optimized for efficient OCR inference and dependable document-processing performance
• Provides a practical balance between memory efficiency, extraction accuracy, and runtime speed
• Suitable for invoices, reports, forms, books, receipts, and enterprise document-processing pipelines
• Maintains reliable performance across a broad range of OCR and document-understanding tasks

---

Training Background (Original Model)

DeepSeek-OCR-2 is trained with a strong emphasis on document parsing, optical character recognition, and multimodal document understanding across diverse visual document collections.

Pretraining

• Large-scale multimodal pretraining using document-centric image and text datasets
• Focus on visual-text alignment, document representation learning, and OCR capability
• Optimized for downstream document understanding and structured extraction tasks

OCR Optimization

• Further optimized for document parsing and structured information extraction
• Enhanced for robust text recognition across diverse layouts and document formats
• Improved performance on tables, forms, receipts, reports, and visually complex documents

---

Key Capabilities

• Optical Character Recognition (OCR) : Accurately extracts textual content from scanned documents, images, and visual media.

• Document Parsing : Understands document layout and semantic organization for structured processing.

• Layout Understanding : Identifies relationships between document elements including paragraphs, tables, and forms.

• Structured Information Extraction : Extracts key-value information and structured content from complex documents.

• Complex Document Analysis : Handles visually rich documents containing mixed layouts, tables, and embedded content.

• Efficient Local Deployment : Quantized variants enable practical OCR inference on consumer hardware.

---

Usage Example

Using llama.cpp

./llama-mtmd-cli \
  -m SandlogicTechnologies/DeepSeek-OCR-2_IQ4_NL.gguf \
  --mmproj SandlogicTechnologies/deepseek-ocr-2.mmproj-f16.gguf \
  --image invoice.png \
  -p "Extract all document fields and return them as structured JSON."

---

Recommended Usecases

• Document Digitization : Convert scanned documents into searchable digital content.

• Enterprise OCR : Automate extraction of structured information from business documents.

• Invoice & Receipt Processing : Capture financial information from invoices and receipts.

• Document Parsing Pipelines : Build automated workflows for layout-aware document understanding.

• Knowledge Extraction : Prepare structured document data for indexing, search, and downstream AI systems.

• Research & Evaluation : Benchmark OCR and document-understanding performance across diverse datasets.

---

Acknowledgments

These quantized models are based on the original work by the DeepSeek AI development team.

Special thanks to:

• The DeepSeek AI team for developing and releasing the DeepSeek-OCR-2 model.
• Georgi Gerganov and the llama.cpp open-source community for enabling efficient quantization and inference via the GGUF format.

---

Contact

For questions, feedback, or support, please reach out at support@sandlogic.com or visit https://www.sandlogic.com/