DeepSeek-OCR: The SOTA Contexts Optical Compression for Document AI
Unlock unprecedented document understanding and high-speed OCR with the DeepSeek-OCR model. From image-to-markdown to complex table parsing, experience next-generation Contexts Optical Compression.
High-Speed Inference up to 2500 tokens/s on an A100-40G.
from 99+ happy users
Get Started with DeepSeek-OCR
Integrate the model into your workflow in three simple steps:
What is DeepSeek-OCR
DeepSeek-OCR is an advanced Multimodal Large Language Model (MLLM) focused on optimizing the Optical Compression of visual data, pushing the boundaries of Visual-Text Compression for complex documents.
- LLM-Centric VisionA novel approach investigating the role of vision encoders from an LLM-centric viewpoint for better context comprehension.
- Contextual OCR & ParsingGoes beyond simple text recognition to understand document structure, layouts, and contextual meaning.
- Scalable Resolution SupportSupports multiple native and dynamic resolutions to handle diverse document sizes without losing fidelity.
Core Capabilities of DeepSeek-OCR
From high-fidelity text extraction to complex document structure conversion, DeepSeek-OCR handles all your Document AI needs.
Prompting for Powerful Document AI
Leverage DeepSeek-OCR's contextual understanding with specialized prompts.
Contextual Document Parsing
Prompt: `<image>\n<|grounding|>Convert the document to markdown.` (Ideal for structured extraction).
Precise Text Localization (REC)
Prompt: `<image>\nLocate <|ref|>xxxx<|/ref|> in the image.` (For finding specific text blocks).
Freeform/General OCR
Prompt: `<image>\nFree OCR.` (Simple, unconstrained text extraction).
Figure and Chart Analysis
Prompt: `<image>\nParse the figure.` (To understand visual elements within the document).
Detailed Image Description
Prompt: `<image>\nDescribe this image in detail.` (General multimodal capability).
Multilingual Support
Tested on various languages, including complex scripts like Chinese.
Unrivaled OCR Speed and Efficiency
DeepSeek-OCR is engineered for speed and high-volume processing.
Inference Speed
~2500
tokens/s (PDF Concurrency on A100-40G)
Resolution Modes
5+
Supported Resolution Modes (Including Dynamic)
GitHub Stars
7.2k+
Community Support
Frequently Asked Questions About DeepSeek-OCR
For more technical details, please refer to the GitHub repository or the paper.
What is 'Contexts Optical Compression'?
It's the core concept of DeepSeek-OCR, focusing on how a Vision Encoder effectively compresses rich visual context into tokens for the LLM, ensuring maximum fidelity for document understanding.
What hardware is required for high-speed inference?
The vLLM inference shows its best performance on NVIDIA A100-40G or similar high-end GPUs, achieving speeds up to 2500 tokens/s, which is critical for large-scale PDF processing.
Does DeepSeek-OCR support structured data extraction?
Yes. By using the `<|grounding|>Convert the document to markdown.` prompt, the model can effectively extract and structure document content, including tables and lists, into a clean format.
What resolution modes are available?
The model supports Tiny (512×512), Small (640×640), Base (1024×1024), Large (1280×1280) native modes, and a dynamic Gundam mode for flexible and high-fidelity inputs.
How can I contribute or report issues?
You can contribute by forking the official DeepSeek-OCR GitHub repository, submitting Pull Requests, or opening new Issues for bugs and suggestions.
Where can I download the model weights?
The official model weights are hosted on the Hugging Face Hub under the `deepseek-ai/DeepSeek-OCR` repository.
Start Building with DeepSeek-OCR Today
Experience the next-generation performance of Document AI and Contexts Optical Compression.