How to Run DeepSeek OCR on Ollama: Local Setup Guide (2026)

What DeepSeek OCR Is and How Optical Compression Works

DeepSeek OCR is built from two pieces: a DeepEncoder with roughly 380 million parameters that combines window attention (based on SAM-base) and global attention (based on CLIP-large) through a 16x convolutional compressor, and a DeepSeek3B-MoE-A570M decoder that turns the compressed vision tokens into text. The whole pipeline totals 3.34 billion parameters, small enough that Ollama distributes it as a single 6.7 GB F16 file rather than the multi-hundred-gigabyte downloads that cloud-only models like MiniMax M3 require.

The detail that makes DeepSeek OCR worth a dedicated guide is optical context compression. A typical page of dense text might run 6,000 or more tokens once it's parsed into raw text. DeepSeek OCR instead encodes the rendered page image into a much smaller number of vision tokens and reconstructs the text from those. On OmniDocBench, DeepSeek's own benchmark, the model matches GOT-OCR2.0's accuracy (which needs 256 tokens per page) using only about 100 vision tokens, and it outperforms MinerU2.0 (which averages over 6,000 tokens per page) while using fewer than 800 vision tokens. On the independent olmOCR-bench benchmark, DeepSeek OCR scores 75.7 overall, with 77.2 on Arxiv Math documents and 73.6 on old, lower-quality scans.

Here's how DeepSeek OCR compares to two of the document-parsing tools it benchmarks against:

The model card on Hugging Face (`deepseek-ai/DeepSeek-OCR`) also documents five resolution presets, Tiny, Small, Base, Large, and Gundam, that trade accuracy for speed by changing how large the input image is and whether it gets cropped into tiles before encoding. Ollama's `deepseek-ocr` tag ships with a fixed default and does not expose these presets as a runtime option, which the next section covers in more detail.

Install Ollama and Run DeepSeek OCR

Step 1: Install Ollama

# Linux and macOS, one-command installer
curl -fsSL https://ollama.com/install.sh | sh

On Windows, download the installer from ollama.com/download, or use winget:

winget install Ollama.Ollama

Step 2: Confirm You're on Ollama 0.13.0 or Later

ollama --version

Step 3: Pull and Run DeepSeek OCR

ollama run deepseek-ocr

This downloads the 6.7 GB model file (tagged `deepseek-ocr:3b`, aliased to `latest`) the first time you run it, then drops you into an interactive prompt. Subsequent runs skip the download.

Step 4: Run OCR on an Image

Pass the image path and an instruction together, separated by a literal `\n`:

ollama run deepseek-ocr "/path/to/invoice.png\nFree OCR."

Expected output is the plain text extracted from the image, printed directly to the terminal. For a scanned invoice, that might look like:

INVOICE #4471
Date: 2026-06-18
Bill To: Acme Corp
Subtotal: $1,240.00
Tax: $99.20
Total: $1,339.20

OCR Prompts: Plain Text, Markdown, Layout, and Figures

DeepSeek OCR responds to a small set of documented prompt patterns, each suited to a different task. All of them follow the same structure: the image path, a literal `\n`, then the instruction.

The `<|grounding|>` tag tells the model to track where text sits on the page, which is what makes the Markdown conversion preserve headings, lists, and table structure instead of flattening everything into one block of text:

ollama run deepseek-ocr "/path/to/report-page-1.png\n<|grounding|>Convert the document to markdown."

Batch-Converting a Folder of Scanned Pages

For multi-page documents exported as individual images, a simple loop processes each page and writes the Markdown output to its own file:

for img in pages/*.png; do
  name=$(basename "$img" .png)
  ollama run deepseek-ocr "$img\n<\|grounding\|>Convert the document to markdown." > "output/$name.md"
done

Concatenate the resulting files afterward if you need a single combined Markdown document.

About the Tiny/Small/Base/Large/Gundam Resolution Presets

DeepSeek's Hugging Face model card documents five resolution presets that change how the image gets resized and whether it's cropped into tiles before encoding, from Tiny (512x512, fastest, lowest accuracy on dense text) up to Gundam (1024 base size with 640px tiled crops, the preset DeepSeek recommends for dense documents). Ollama's `deepseek-ocr` tag does not expose these presets through the command line or API; it runs with a single fixed default. If you specifically need control over resolution mode, for example to push accuracy higher on a dense academic paper, that requires running the model directly through Hugging Face transformers rather than through Ollama.

Use DeepSeek OCR From Scripts via the Ollama API

Ollama exposes every local model, including `deepseek-ocr`, through its REST API on `localhost:11434`, which means you can build an OCR pipeline without shelling out to the CLI for every image.

Call DeepSeek OCR with curl

Ollama's API takes images as base64-encoded strings in the `images` array:

IMAGE_B64=$(base64 -w0 invoice.png)

curl http://localhost:11434/api/generate -d "{
  \"model\": \"deepseek-ocr\",
  \"prompt\": \"Free OCR.\",
  \"images\": [\"$IMAGE_B64\"],
  \"stream\": false
}"

Expected output (truncated):

{
  "model": "deepseek-ocr",
  "response": "INVOICE #4471\nDate: 2026-06-18\nBill To: Acme Corp\n...",
  "done": true
}

Python Example

import base64
from ollama import Client

client = Client(host="http://localhost:11434")

with open("invoice.png", "rb") as f:
    image_b64 = base64.b64encode(f.read()).decode()

response = client.generate(
    model="deepseek-ocr",
    prompt="<|grounding|>Convert the document to markdown.",
    images=[image_b64],
)
print(response["response"])

This is the same pattern covered in the Ollama with Python guide, with an `images` argument added for the vision input.

Troubleshooting

Frequently Asked Questions

ollama run deepseek-ocr "/path/to/page.png\n<|grounding|>Convert the document to markdown."