# DeepSeek OCR: what it is, what it needs, and how to use it without a GPU

DeepSeek-OCR is an open-source vision model that DeepSeek released on October 20, 2025, and you run it yourself on an NVIDIA GPU with CUDA 11.8 and PyTorch 2.6[1]. It turns a page image into Markdown. If you want a scan or photo as text you can edit, LoreAI reads it with Mistral OCR in the browser and returns a Word file.

## What is DeepSeek-OCR?

DeepSeek-OCR is a research model from the Chinese AI lab DeepSeek, published with the paper DeepSeek-OCR: Contexts Optical Compression[2]. It has two parts, an image encoder called DeepEncoder and a 3B mixture-of-experts language model as the decoder[2]. The idea behind it is to squeeze a page of text into a small number of vision tokens, which is why the launch drew so much attention from people who build AI systems.

The weights are on Hugging Face under the MIT license, so anyone can download and run them[3]. DeepSeek followed it with DeepSeek-OCR 2 on January 27, 2026, with the same CUDA 11.8 and PyTorch 2.6 setup[4].

## DeepSeek-OCR at a glance

- **Released**: October 20, 2025, and DeepSeek-OCR 2 on January 27, 2026[1]
- **License**: MIT for DeepSeek-OCR[3], Apache 2.0 for DeepSeek-OCR 2[5]
- **Runs on**: An NVIDIA GPU, tested with CUDA 11.8, PyTorch 2.6 and Python 3.12.9[3]
- **Modes**: Tiny 512×512, Small 640×640, Base 1024×1024, Large 1280×1280, and Gundam for large pages[1]
- **Output**: Markdown, with the prompt Convert the document to markdown[1]
- **PDF**: A separate script processes PDF files[1]
- **Languages**: Nearly 100 languages for PDF documents, according to the paper[6]

## How to use DeepSeek OCR

DeepSeek-OCR is a model you install on a computer with a suitable graphics card. The GitHub repository describes two routes, vLLM for speed and Hugging Face Transformers for a quick test[1].

1. **Set up the environment.** Install Python 3.12.9, PyTorch 2.6.0 and the vLLM or Transformers versions the repository lists, on a machine with an NVIDIA GPU and CUDA 11.8[1].
2. **Download the model.** Load deepseek-ai/DeepSeek-OCR from Hugging Face[3].
3. **Choose a mode and a prompt.** Base or Gundam suits a full page. The prompt Convert the document to markdown keeps the layout, and Free OCR returns plain text without it[1].
4. **Run the image or PDF script.** The repository has one script for images and one for PDFs[1]. The result is Markdown you open in a text editor.

DeepSeek reports about 2,500 tokens a second for PDFs on an A100 40 GB card[1], a data-center GPU. A student laptop has no such card.

## Is there a DeepSeek OCR online demo?

Not from DeepSeek. The GitHub repository and the model card link no hosted demo[1][3]. The DeepSeek OCR demo pages that rank in search are Hugging Face Spaces and websites built by other people around the open weights. Check who runs a site before you upload a document to it.

## Can DeepSeek read images?

Through the API, yes. DeepSeek's API documentation says the deepseek-flash model accepts images with text, and lists reading text from screenshots among its uses, for JPEG, PNG, GIF and WebP files[7]. Each image can be up to 32 MiB when sent as base64 or a link[7]. That is a chat model answering about an image, which is a different tool from the DeepSeek-OCR model.

For a page you want back as a document, a chat answer leaves you to copy, paste and rebuild the tables yourself. The same is true of [ChatGPT OCR](https://www.loreai.org/compare/chatgpt-ocr) and [Gemini OCR](https://www.loreai.org/compare/gemini-ocr).

## DeepSeek-OCR and LoreAI side by side

*Checked on November 10, 2026, against DeepSeek's GitHub, Hugging Face and API pages*

| What you need | DeepSeek-OCR | LoreAI |
| --- | --- | --- |
| File types | Images, plus PDF through a script[1] | PDF, JPG, PNG and HEIC |
| Size | Not stated in the repository | Free 3 MB, Basic 10 MB, Pro 25 MB |
| Length | Not stated in the repository | Free 3, Basic 30, Pro 100 pages in one file |
| Scanned pages | Read as page images[1] | Every page read with Mistral OCR, in order |
| Photos | Read as images[1] | Uploaded from the phone, HEIC included |
| What you download | Markdown text[1] | An editable Word file (.docx) with headings, lists and tables |
| Phone | No, it runs on a GPU machine[3] | Any phone browser |
| Translation | Not part of the model's prompts[1] | Into 105 languages in the same step |
| Price | Free weights under the MIT license[3], you supply the GPU | First document free, then Basic at $12.99 a month |

## Which one should you use?

- Use DeepSeek-OCR if you build a document pipeline, have GPU time and want an open model you control.
- Use the DeepSeek API if you want to ask questions about a screenshot inside your own app.
- Use LoreAI if you have a scan, a phone photo or a stack of handwritten notes and want an editable Word file in a minute, with no install.

LoreAI uses Mistral OCR, not DeepSeek. The [Mistral OCR comparison](https://www.loreai.org/compare/mistral-ocr) explains how that model is used, and the [OCR software comparison](https://www.loreai.org/compare/best-ocr-software) covers the desktop tools.

## Questions people ask

### Is DeepSeek OCR free?

The model weights are free under the MIT license[3]. You pay for the GPU that runs it.

### Can I use DeepSeek OCR online?

DeepSeek does not host a demo for it. Its repository and model card link none[1][3]. LoreAI reads scans in the browser with Mistral OCR.

### Can DeepSeek read images?

Yes, through the API. The deepseek-flash model accepts JPEG, PNG, GIF and WebP images and can read text from screenshots[7].

### What output does DeepSeek-OCR give?

Markdown, when you use the prompt Convert the document to markdown, or plain text with the prompt Free OCR[1].

### What is DeepSeek-OCR 2?

The second version, released on January 27, 2026, under the Apache 2.0 license[4][5].

## Sources

1. [GitHub: deepseek-ai/DeepSeek-OCR, README](https://github.com/deepseek-ai/DeepSeek-OCR), accessed November 10, 2026
2. [arXiv 2510.18234: DeepSeek-OCR: Contexts Optical Compression](https://arxiv.org/abs/2510.18234), accessed November 10, 2026
3. [Hugging Face: deepseek-ai/DeepSeek-OCR model card](https://huggingface.co/deepseek-ai/DeepSeek-OCR), accessed November 10, 2026
4. [GitHub: deepseek-ai/DeepSeek-OCR-2, README](https://github.com/deepseek-ai/DeepSeek-OCR-2), accessed November 10, 2026
5. [Hugging Face: deepseek-ai/DeepSeek-OCR-2 model card](https://huggingface.co/deepseek-ai/DeepSeek-OCR-2), accessed November 10, 2026
6. [arXiv 2510.18234, full text (HTML)](https://arxiv.org/html/2510.18234), accessed November 10, 2026
7. [DeepSeek API Docs: Vision](https://api-docs.deepseek.com/guides/vision), accessed November 10, 2026

By Matt Mountain · Updated November 10, 2026