Skip to content
Get packs

DeepSeek-OCR

An OCR model that turns document images into text and markdown, from DeepSeek AI.

GitHub Download source
Stars
24.1k
Forks
2.2k
Licence
MIT
Language
Python
Last update
9 months ago
Type
AI model / agent

What it is

DeepSeek-OCR is a research model released by DeepSeek AI that reads text from images, such as scanned pages or PDFs, and converts them into text or markdown. The README frames it as an investigation into compressing text into visual tokens, using vision encoders from an LLM-centric viewpoint. It is mainly used by developers who want to run OCR locally through Python.

  • Convert a document image to markdown using the grounding prompt shown in the README
  • Run plain text extraction on an image with the 'Free OCR.' prompt
  • Process PDFs in batches with the included run_dpsk_ocr_pdf.py script
  • Run batch evaluation for benchmarks with run_dpsk_ocr_eval_batch.py

Great for

  • Developers who want an open-weight OCR model they can run themselves
  • Teams converting scanned documents into markdown or plain text
  • Researchers studying visual-text compression with vision encoders
  • Users who already have a CUDA GPU environment set up

Not ideal for

  • People who need a point-and-click app: the README only shows Python scripts and commands
  • Users without an NVIDIA/CUDA GPU setup, since the install instructions assume CUDA 11.8 and torch 2.6.0
  • Teams needing guaranteed accuracy on specific document types, since the README shows examples but no accuracy benchmarks in the truncated text
Step-by-step guide: how to use DeepSeek-OCRInstall, first run, examples, tips and FAQs
DeepSeek-OCR vs Unlimited OCR

How to get it

  1. 1

    Clone and build from source

    You'll need: Python 3.12.9 (via conda, per README) · CUDA 11.8 with torch 2.6.0 (per README) · Conda

    Clone the repository

    git clone https://github.com/deepseek-ai/DeepSeek-OCR.git

    Create and activate the conda environment

    conda create -n deepseek-ocr python=3.12.9 -y && conda activate deepseek-ocr

    Install PyTorch (CUDA 11.8 build)

    pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu118

    Install project requirements

    pip install -r requirements.txt
  2. 2

    Run it

    • Set the INPUT_PATH and OUTPUT_PATH in DeepSeek-OCR-master/DeepSeek-OCR-vllm/config.py
    • Change into the DeepSeek-OCR-master/DeepSeek-OCR-vllm folder
    • Run python run_dpsk_ocr_image.py to process an image with streaming output
  3. 3

    Get the most out of it

    • Use the grounding prompt '<image>\n<|grounding|>Convert the document to markdown.' for document-to-markdown output
    • Use 'Free OCR.' when you only need plain text without layout
    • Pick a resolution mode from the Support-Modes list to trade detail against vision token count
    • Use the vLLM path for batched PDF processing, as the README reports about 2500 tokens/s on an A100-40G

Downloads

Source code (.zip)main branch · straight from GitHub

Strengths

  • Open weights and MIT-licensed code, with a Hugging Face model download
  • Supports both vLLM and Transformers inference paths in the README
  • Offers several resolution modes (Tiny to Large, plus Gundam) for different speed and detail trade-offs

Trade-offs

  • Setup is involved: it requires a specific CUDA/torch version, a vLLM wheel, and flash-attn
  • The README is truncated and does not state hardware requirements beyond an A100-40G throughput note

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.