DeepSeek OCR, online

What is DeepSeek-OCR — and why Unlimited-OCR is better

DeepSeek-OCR is a vision-language OCR model that reads dense pages by compressing them through its DeepEncoder vision stack (SAM + CLIP) and decoding the result straight into structured text.

Unlimited-OCR (DeepSeek-based) is its improved successor, built directly on the DeepSeek-OCR DeepEncoder architecture. It keeps the same DeepEncoder but pairs it with a 3B Mixture-of-Experts decoder and a constant-size KV cache, so it parses long, multi-page documents in one shot without memory blowing up — cleaner, more reliable extraction on exactly the documents DeepSeek-OCR targets.

Run it online here — nothing to install

No Python, CUDA, or GPU server setup. chunks.md runs the INT4 model fully in your browser with WebGPU, so the vision and decoder work happens on your own device. Drop in a PDF or image and get clean markdown back — layout-aware and multilingual, including Japanese (for manga pages specifically, manga-ocr is the better fit). Files never leave your machine, and the model caches after the first load.

The markdown output is built for reuse — feed it straight into AI assistants and knowledge bases. And if you just need fast text from simple documents, PaddleOCR is the lighter, quicker alternative.

How to run DeepSeek OCR online

1. Open chunks.md and pick Unlimited-OCR (DeepSeek-based) from the model list.
2. Drop in a PDF or image — research papers, invoices, multi-column scans, dense tables.
3. Press Chunk it. The model runs on your device via WebGPU; the first run downloads and caches it.
4. Copy the layout-aware markdown, or download it as a .md file.

No account, no upload, no server queue — the document never leaves your machine.

Related reading