# DeepSeek-OCR: Contexts Optical Compression

**Authors:** Haoran Wei, Yaofeng Sun, Yukun Li  
**arXiv:** [2510.18234](https://arxiv.org/abs/2510.18234)  
**Date:** 2025-10-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2510.18234v1/

## Key points

- DeepSeek-OCR's deep parsing abilities allow it to analyze images within documents through secondary model calls
- The model can extract structured information from various types of images such as charts, natural images, chemical formulas, and geometric figures with just one unified prompt
- DeepSeek-OCR showcases versatility by performing deep parsing on financial charts, natural images, chemical formulas, and planar geometric figures
- It demonstrates impressive multilingual recognition proficiency by handling nearly 100 languages present in PDF documents on the internet
- The adaptability of DeepSeek-OCR to different languages emphasizes its utility in processing multilingual data for LLM/VLM pretraining
- Its practical performance is highlighted by generating training data at scale for LLMs/VLMs with high OCR accuracy compared to existing models like GOT-OCR2.0 and MinerU2.0
- Overall, DeepSeek-OCR's capabilities make it a promising tool for research areas such as historical long-context compression and memory forgetting mechanisms in LLMs

## Summary

, , , , 

In a recent study, the capabilities of DeepSeek-OCR were explored for its deep parsing abilities. This allows the model to analyze images within documents through secondary model calls. With this feature, known as "deep parsing," the model can extract structured information from various types of images such as charts, natural images, chemical formulas, and geometric figures with just one unified prompt. The versatility of DeepSeek-OCR is showcased through its ability to perform deep parsing on financial charts, natural images in books and articles, chemical formulas in STEM documents, and planar geometric figures. This highlights its potential applications in diverse fields. Furthermore, DeepSeek-OCR demonstrates impressive multilingual recognition proficiency by handling nearly 100 languages present in PDF documents on the internet. This capability is crucial for training Large Language Models (LLMs) as it supports both layout and non-layout OCR formats for languages like Arabic and Sinhala. The adaptability of DeepSeek-OCR to different languages further emphasizes its utility in processing multilingual data for LLM/VLM pretraining. Additionally, its practical performance is highlighted by its ability to generate training data at scale for LLMs/VLMs. By achieving high OCR accuracy with minimal vision tokens on OmniDocBench compared to existing models like GOT-OCR2.0 and MinerU2.0, DeepSeek-OCR proves its efficiency in producing quality training data efficiently. Overall,<Organization>'s deep parsing capabilities, multilingual recognition proficiency, and practical performance make it a promising tool for various research areas such as historical long-context compression and memory forgetting mechanisms in LLMs. Its accessibility through publicly available codes and model weights further enhances its value for researchers seeking advanced optical mapping solutions.