DeepSeek-OCR – A Vision-Language Model Open-Sourced by the DeepSeek Team

What is DeepSeek-OCR？

DeepSeek-OCR is a vision-language model developed by the DeepSeek team, designed to efficiently process long-text content through optical compression technology. The model consists of the DeepEncoder encoder and the DeepSeek-3B-MoE decoder, achieving high-resolution input processing while significantly reducing activation memory and the number of visual tokens. DeepSeek-OCR achieves 97% OCR accuracy at a 10× compression ratio and maintains 60% accuracy at 20× compression. It supports multiple resolution modes, multilingual document processing, and complex content parsing (e.g., charts and chemical formulas), providing a highly efficient solution for large-scale document processing.

Key Features of DeepSeek-OCR

• Visual Text Compression: Compresses long-text content efficiently through visual modalities, achieving a 7–20× compression ratio.

• Multilingual OCR: Supports document recognition in nearly 100 languages, including Chinese, English, Arabic, and Sinhala.

• Deep Parsing: Accurately interprets charts, chemical formulas, and geometric figures.

• Multi-Format Output: Supports both layout-preserving Markdown and plain OCR output formats.

Technical Architecture of DeepSeek-OCR

Core Structure

DeepSeek-OCR is composed of two main components:

• DeepEncoder – Encodes input images (documents) into visual tokens.

• DeepSeek-3B-MoE-A570M – Serves as the decoder, converting visual tokens into text.

DeepEncoder

The DeepEncoder is the core component of DeepSeek-OCR, designed to maintain low activation memory under high-resolution input while achieving high compression. It includes the following components:

• Dual-Tower Structure:

  • SAM-base (80M): Utilizes window attention to capture local features with low memory usage for high-resolution input.

  • CLIP-large (300M): Uses global attention to extract semantic information. Since the input is compressed, memory usage remains efficiently controlled.

• 16× Convolutional Compression Layer:
  Between SAM and CLIP, a 16× convolutional compression module reduces the number of visual tokens from 4096 to 256 through two convolutional layers (stride=2). This design minimizes memory usage while preserving key information.

• Multi-Resolution Support:
  DeepEncoder supports multiple resolution modes — Tiny, Small, Base, Large, and Gundam — each corresponding to different input resolutions and token counts:

  • Tiny: 512×512 resolution → 64 visual tokens

  • Small: 640×640 resolution → 100 visual tokens

  • Base: 1024×1024 resolution → 256 visual tokens

  • Large: 1280×1280 resolution → 400 visual tokens

  • Gundam: Dynamic resolution mode for even higher input resolutions, with block-based processing to further reduce activation memory.

Decoder: DeepSeek-3B-MoE-A570M

The decoder is based on the DeepSeek-3B-MoE architecture with 570M active parameters. It decodes compressed visual tokens into text via nonlinear mapping:

fdec:Rn×dlatent→RN×dtextfdec​:Rn×dlatent​→RN×dtext​

where n is the number of visual tokens, N is the number of text tokens, and dₗₐₜₑₙₜ and dₜₑₓₜ represent the dimensions of the visual and text tokens respectively.

Project Links

• GitHub Repository: https://github.com/deepseek-ai/DeepSeek-OCR

• Hugging Face Model Hub: https://huggingface.co/deepseek-ai/DeepSeek-OCR

• Technical Paper: DeepSeek_OCR_paper.pdf

Application Scenarios

• Large-Scale Training Data Generation: Automatically processes hundreds of thousands of document pages daily, generating high-quality datasets for large language and vision-language models.

• Enterprise Document Digitization: Rapidly converts paper-based contracts, reports, and other corporate documents into searchable, editable digital formats.

• Academic Research and Literature Processing: Accurately parses complex academic content such as mathematical equations, chemical formulas, and figures into structured, machine-readable data.

• Multilingual Global Document Handling: Efficiently manages multilingual document workflows for multinational enterprises and global organizations.

• Financial and Business Intelligence Analysis: Deeply analyzes research report charts and converts them into structured data to support automated financial analysis and investment decision-making.