Reads text, tables and formulas out of scanned pages and phone photos, even when the page is curved

You give it an image of a page and it gives you back the content as structured text. Plain paragraphs come out as text. Tables come out as real tables with rows, columns, and merged cells preserved. Formulas come out as LaTeX. Code blocks come out as code. Ask it for the layout and it marks where each block sits on the page and in what order to read them.

The distinctive part is the second kind of input. Most parsers expect a straight, evenly lit page. TeleOCR was trained on camera-captured documents as well: pages that are curved, tilted, or folded. It reads the layout directly from the distorted photo, so you do not have to run a dewarping tool first. On the public benchmark for such photos it is currently the top open model, and on the standard clean-document benchmark it also edges out the models we have covered before, including MinerU 2.5 Pro, PaddleOCR-VL 1.6 and FireRed-OCR.

It is a small vision-language model at about 1.2 billion parameters, built on Qwen2.5-VL, and it reads Chinese and English. It was released in August under the name NaviDC-OCR and renamed to TeleOCR this month, so you will see both names around.

The TeleOCR demo Space on Hugging Face is the quick test. Take a photo of a printed page with your phone, holding it at a slight angle so the page bends, upload it, and pick a task: full text, table, formula, or layout. Then upload a clean screenshot of the same kind of page and compare. For your own machine, the model page has a short Python example that runs with transformers on one GPU, and the GitHub repo has the full document-to-Markdown pipeline. There is also a community GGUF build for llama.cpp if you want it on CPU.

Language coverage is Chinese and English only, so pages in other languages are not the target. The transformers example handles one task per prompt; the full parse that stitches layout, text and tables together lives in the GitHub repo, not the model page.

• Turn photos of receipts, forms, or printed pages into editable text.
• Extract a table out of a report photo and paste it into a spreadsheet.
• Convert a textbook page with equations into Markdown with proper LaTeX.
• Digitise a stack of scanned Chinese or English documents in one pass.
• Read the layout of a document to pull just the section you need.

Try the demo

View model page

Read this issue on neuralexpedition.com →