Browser-local processing / Free to use

Convert a PDF to Markdown Locally Beta

Convert PDF pages into Markdown with local layout analysis, OCR, and table recognition. Preserve pictures and merge matching cross-page tables. Your PDF is never uploaded. This feature is in beta; conversion quality and browser compatibility are still being improved.

Source PDF

WebGPU: Checking…

Choose a PDF or drag and drop

PDF • Up to 20 MB and 50 pages • File stays on your device

Select PDF

On-Device Pipeline

Conversion preset
Advanced model settings

17 region classes. Reading order is estimated from region positions; check complex multi-column pages. Only the selected layout model is downloaded. Changing models applies to the next conversion.

FP32 vision encoder 374.1 MB; FP32 token embeddings 231.5 MB; INT8 decoder 164.9 MB; Tokenizer 7.2 MB. Only the selected model downloads, when a formula region needs recognition. Changing models applies to the next conversion.

Encoder 95.3 MB; Structure decoder 77.9 MB; Cell box decoder 39.6 MB. All options use PP-OCR for cell text. Only the selected structure model downloads when a table needs recognition. Changing models applies to the next conversion.

Checking WebGPU support…

Docling Heron INT8 + PP-OCR + TableFormer Accurate + CodeFormulaV2 (INT8 decoder)

Docling Heron INT8 layout weights download about 69 MB, plus 31 MB for OCR and 212.8 MB for table recognition as needed (about 313 MB total). Formula regions additionally download 777.7 MB for CodeFormulaV2 (INT8 decoder) (about 1091 MB total with all models). Cached weights are reused when storage permits. Formulas become editable LaTeX; pictures and failed regions are kept as images. Headers and footers are skipped.

Markdown Result

Your Markdown will appear here

Choose a PDF, then convert. Pages are added here as they finish, so you can copy or download a partial result at any time.

Input tip: WebGPU is required. The first run downloads layout and OCR models.

Share

Documentation

Is WebGPU Mostly Available Now?

Check WebGPU support in Chrome, Safari, Edge, and Firefox, understand platform limits, and learn what browser support means for local OCR.

PDF to Markdown without uploading the file

Accurate is the default preset: Docling Heron INT8 + CodeFormulaV2 (INT8 decoder) + TableFormer Accurate. Fast uses Docling Heron INT8 + Pix2Text-MFR-1.5 FP32 + SLANet+. Expand Advanced model settings to choose individual models, including TableFormer Fast. Choose PP-DocLayoutV3 FP16, Docling Heron INT8, or Docling Egret Medium FP32 to detect page regions inside your browser. PP-DocLayoutV3 predicts reading order; Heron and Egret use region positions to estimate it, so complex columns need extra review. PP-OCR recognizes text, your choice of SLANet+, TableFormer Fast, or TableFormer Accurate recognizes table structure, and your choice of TexTeller3 Q4, Pix2Text-MFR-1.5, or CodeFormulaV2 with an INT8 decoder converts formulas to editable LaTeX. Headers and footers are skipped. Pictures, charts, and regions that fail recognition are preserved as cropped images with visible warnings where appropriate. Pages are converted one at a time and streamed into the editor as they finish, which means you can stop a long job and keep the pages that already completed. Choose a PDF of 20 MB or less and 50 pages or fewer. The file stays on this device throughout, which makes the tool suitable for contracts, reports, and drafts you would rather not send to a third-party API.

What downloads when I convert a PDF?

Only the selected layout model downloads from pinned Hugging Face revisions: about 67 MB for PP-DocLayoutV3 FP16, 69 MB for Docling Heron INT8, or 78 MB for Docling Egret Medium FP32. OCR models add about 31 MB from the Paddle CDN. Table regions download only the selected structure model: about 8 MB for SLANet+, 145.5 MB for TableFormer Fast (78.4 MB encoder, 27.4 MB structure decoder, 39.6 MB cell box decoder), or 212.8 MB for TableFormer Accurate (95.3 MB encoder, 77.9 MB structure decoder, 39.6 MB cell box decoder). All match PP-OCR text to predicted cells. Formula regions download only the selected formula model: TexTeller3 Q4 totals 256.8 MB (56.8 MB encoder, 198.6 MB decoder, 1.4 MB tokenizer), or Pix2Text-MFR-1.5 FP32 totals 119.7 MB (87.5 MB encoder, 32.0 MB decoder, 0.1 MB tokenizer). CodeFormulaV2 totals 777.7 MB: 374.1 MB FP32 vision encoder, 231.5 MB FP32 token embeddings, 164.9 MB INT8 decoder, and 7.2 MB tokenizer. All formula models download directly from pinned Hugging Face revisions. The pipeline panel shows the combined download size for your selections. Weights are cached when browser storage permits. Your PDF is not uploaded, and no page content is sent anywhere. If you would rather not download a model at all, use image to Markdown instead, which recognizes a single exported page image on our server.

Do I need WebGPU?

WebGPU is required for all layout models; there is no CPU fallback. PP-DocLayoutV3 additionally requires FP16 (shader-f16) support. Use Chrome or Edge with GPU acceleration enabled. Close other heavy tabs to free memory before converting.

Which PDFs work best?

Clear scans and readable report pages work best. Every page is rendered as an image for recognition, including PDFs with selectable text. Output can include Markdown and HTML tables. Complex tables, merged cells, formulas, and multiple columns may need manual correction. Check numbers against the original before using the result.

What can I do with the Markdown?

Edit the result, switch to Preview to check tables and preserved images, then copy it or download a .md file. Copy and Download .md embed preserved images as data URLs so the file is self-contained; support for embedded images and HTML tables varies between Markdown readers. Choose Download .docx to export the current edited result as a Word document with editable text, tables, and equations, plus embedded pictures. Both downloads run locally and use the selected cross-page table merge setting. If your source is a set of images rather than a PDF, combine them first with image to PDF and convert the result here, or process a single image with image to Markdown.

How are tables spanning pages handled?

Tables are automatically joined only when they are the last content region of one page and the first of the next, lie near the page boundary, and have matching horizontal positions, column counts, and repeated headers. Repeated headers are removed from the continuation. This conservative check can miss valid continuations. Uncheck Merge matching tables across pages to restore the original tables. Edits are kept separately for the merged and original versions; switching back restores your edits.