llm.create_chat_completion(
messages = "\"cats.jpg\""
)DOTS.OCR
DOTS.OCR is a vision-language document understanding model developed by RedNote HiLab, designed for OCR, document parsing, layout reasoning, and structured content understanding. This repository contains GGUF quantized variants of the model optimized for efficient local inference using llama.cpp.
Rather than functioning as a conventional OCR engine, DOTS.OCR is designed to understand complete document layouts while recognizing textual content. The model jointly interprets paragraphs, tables, mathematical expressions, figures, diagrams, forms, and hierarchical document structures, enabling accurate reconstruction of visually complex documents for downstream AI workflows.
The quantized formats significantly reduce memory requirements while preserving document understanding capability, making the model practical for local deployment, enterprise document intelligence, and large-scale document processing applications.
Model Overview
- Model Name: DOTS.OCR
- Base Model: rednote-hilab/dots.ocr
- Architecture: Vision-Language Document Understanding Model
- Parameter Count: Approximately 1.7 Billion Parameters
- Modalities: Text, Image
- Primary Languages: Multilingual
- Developer: RedNote HiLab
- License: Apache 2.0
Quantization Formats
This repository provides various GGUF quantized versions of the DOTS.OCR model optimized for efficient local inference using llama.cpp.
IQ3_M
- Size reduction of approx 75.41% (836 MB)compared to 16-bit (3.32 GB)
- Aggressive 3-bit quantization optimized for lightweight document understanding and parsing workloads
- Suitable for large-scale document processing on resource-constrained systems
- Enables efficient interpretation of documents containing mixed layouts, tables, figures, and textual content
- Fine-grained preservation of complex layouts, mathematical notation, and structural relationships may decrease compared to higher-precision variants
IQ4_NL
- Size reduction of approx 69.92% (1023 MB) compared to 16-bit (3.32 GB)
- Advanced 4-bit non-linear quantization designed to retain document parsing quality and structural understanding
- Better suited for workflows involving layout-aware analysis, structured extraction, and semantic document interpretation
- Provides improved consistency when processing visually complex documents while minimizing quantization loss
- May require slightly increased computational overhead during inference
IQ4_XS
- Size reduction of approx 71.23% (979 MB) compared to 16-bit (3.32 GB)
- Balanced 4-bit quantization optimized for efficient document reasoning and dependable parsing performance
- Provides a practical balance between memory efficiency, document comprehension quality, and runtime speed
- Suitable for document intelligence systems, content indexing, document conversion, and structured information extraction
- Maintains stable performance across diverse real-world document formats
Training Background (Original Model)
DOTS.OCR is trained with an emphasis on document understanding, multimodal reasoning, layout analysis, and structured document interpretation across a diverse collection of visual documents.
Pretraining
- Large-scale multimodal pretraining using document-centric image and text datasets
- Focus on visual-text alignment, semantic document representation, and layout-aware learning
- Optimized for downstream OCR, document parsing, and structured information understanding
Document Understanding Optimization
- Enhanced for preserving document hierarchy, reading order, and structural semantics
- Optimized for interpreting tables, forms, equations, diagrams, and mixed-content layouts
- Improved consistency for document parsing and machine-readable document reconstruction
Key Capabilities
Document Parsing Understands complete document layouts while preserving structural relationships.
Optical Character Recognition (OCR) Extracts textual information from scanned documents and visual content.
Layout Understanding Identifies logical organization across headings, paragraphs, tables, figures, and forms.
Structured Content Extraction Generates machine-readable representations while maintaining semantic document structure.
Complex Document Analysis Processes visually rich documents containing mixed layouts, equations, diagrams, and technical content.
Efficient Local Deployment Quantized variants enable practical document intelligence workloads on consumer hardware.
Usage Example
Using llama.cpp
./llama-mtmd-cli \
-m SandlogicTechnologies/DOTS-OCR_IQ4_NL.gguf \
--mmproj SandlogicTechnologies/mmproj-dots.ocr-f16.gguf \
--image research_paper.png \
-p "Convert this document into structured Markdown while preserving headings, tables, equations, and figures."
Recommended Usecases
Document Parsing Convert visually complex documents into structured machine-readable formats.
Enterprise Document Intelligence Automate understanding of reports, contracts, manuals, and technical documentation.
Layout-Aware Information Extraction Preserve document hierarchy and semantic relationships during extraction.
Knowledge Base Construction Prepare structured documents for enterprise search and Retrieval-Augmented Generation (RAG).
Technical Document Processing Parse scientific papers, books, forms, diagrams, and engineering documentation.
Research and Evaluation Benchmark multimodal document understanding and layout reasoning capabilities.
Acknowledgments
These quantized models are based on the original work by the RedNote HiLab development team.
Special thanks to:
- The RedNote HiLab team for developing and releasing the DOTS.OCR model.
- Georgi Gerganov and the
llama.cppopen-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/
- Downloads last month
- -
3-bit
4-bit
Model tree for SandLogicTechnologies/dots.mocr-GGUF
Base model
rednote-hilab/dots.ocr
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="SandLogicTechnologies/dots.mocr-GGUF", filename="", )