What is chunks.md?
Eight OCR models — PaddleOCR, SmolDocling, MangaOCR, DeepSeek-based Unlimited-OCR — running directly in your browser. One site, every model, zero uploads. Your device does the work.
Documents in, markdown out
Drop a PDF, photo, or scan into chunks.md and get structured markdown back. Text, headings, tables, code blocks — extracted and formatted, ready to paste into your workflow.
Everything runs in your browser. Your files never leave your device. No account required, no upload limits, no data collection.
Eight models, one interface
The latest PaddleOCR model. Highest accuracy across 11 language packs. Best for most documents. 92 MB.
PaddleOCR v5 Mobile
Balanced accuracy at just 12 MB. The default on mobile devices.
PaddleOCR v3
Lightweight and fast at ~1 s/page. Just 11 MB. Good for quick extractions when speed matters more than precision.
SmolDocling 256M
IBM vision-language model that understands document layout. Outputs structured markdown with tables, formulas, and headings. English-only.
PaddleOCR-VL 1.5
Vision-language model combining NaViT encoder with ERNIE decoder. Understands visual context, tables, and page structure. Customizable prompts. INT8 and FP32 variants.
Unlimited-OCR (DeepSeek-based)
DeepEncoder-architecture vision model with WebGPU acceleration. Layout-aware, multilingual including Japanese. Needs a WebGPU-capable device.
TrOCR-based model specialized for Japanese text in manga and print. Handles vertical and horizontal text, furigana, and varied font styles. INT8 and FP32 variants.
PDF Embed Text
Instant extraction of text already embedded in a PDF — no OCR pass at all. The fastest path for digital-native PDFs.
Formats & languages
PDF, JPG, PNG, and WebP. Multi-page PDFs are processed page-by-page with real-time progress.
10+ language groups: English, Chinese, Japanese, Korean, Arabic, Hindi, Latin scripts, Cyrillic, Thai, Greek, Tamil, and Telugu.
Works offline
Models are cached locally in your browser after first use. Come back anytime — chunks.md works without an internet connection from your second visit onward.
One site to access them all
No hunting for separate tools. PaddleOCR, SmolDocling, MangaOCR — every model in a single tab. Switch models instantly and compare results side-by-side.
From quick text extraction to full document understanding to Japanese manga — every model runs locally, privately, and for free. The markdown you get back is ready to feed into AI assistants as clean, structured knowledge.
Related reading
Umi-OCR online
Want Umi-OCR without the install? Run the same PaddleOCR engine online, free — OCR happens on your device, output is clean markdown.
Manga OCR online
Run the manga-ocr model in your browser, free. Extract Japanese text from manga panels and full pages — vertical text, furigana, no uploads, no install.
DeepSeek OCR online
Try DeepSeek-style OCR online: Unlimited-OCR (DeepSeek-based) runs in your browser with WebGPU. Layout-aware, multilingual, markdown out — no upload.