Subscribe
TEST RECORD · ← all field reports

FIELD REPORTS/FIELD NOTE/ISSUE #132

Turn scanned pages into clean text, key fields, and figure captions

Getting the words off a scan is the easy part. The hard part is the invoice total, the table that runs onto the next page, the chart nobody ever typed up. This small model takes on those jobs too, and you can try it in the browser.

MODELOvisOCR3
PUBLISHEDOctober 27, 2026
READ TIME3 min
CATEGORYFIELD NOTE

Field notes

01 What it does

You give it a page image or a PDF and pick a job. It can turn the page into clean Markdown with tables and formulas intact, list every word with its position, or answer a question about the page. For forms, receipts and invoices, you name the fields you want, say invoice number, date and total, or paste a JSON template, and it fills them in. It also describes the charts and diagrams on a page in a few sentences and turns flowcharts into editable Mermaid code. When a table breaks across two pages, it stitches the two halves back into one.

It is small for what it does: 0.8 billion parameters, built on Qwen3.5-0.8B. In the developers' benchmarks it beat other compact OCR models such as PaddleOCR-VL and GLM-OCR on most document tests, and it scored highest of every model compared at pulling key fields out of forms. It can also translate a selected part of the page into another language.

It comes from the Ovis team at Alibaba's AIDC-AI and is released under Apache 2.0, so you can use it in commercial work.

02 How to try it

Open the OvisOCR3 Space. Upload an image or a PDF of up to 100 pages, choose a task and run it. Start with a photo of a receipt or invoice and key information extraction: ask for the vendor, date and total, then check the result against the paper. Next, drop in a report page with a chart and run figure captioning to see how it describes what the chart shows. If you have a PDF where a table spills onto the next page, give it those two pages and run cross-page table reconstruction. The demo is public, so stick to documents you would not mind sharing. To run it yourself, the model card has a short vLLM script and ready-made prompt files for each task. You will need a graphics card.

03 Caveat

It can still miss or repeat text on small, blurry or skewed pages and very dense layouts, and its chart descriptions can add details that are not on the page. Check the output against the source before you rely on it.

04 What you can do with it

  • Turn a stack of scanned receipts into a spreadsheet of dates and totals.
  • Convert old PDFs into Markdown you can edit, search or hand to another AI tool.
  • Pull the figures out of research papers with a short description of each chart.
  • Rebuild tables that break across pages in financial reports.
  • Read the text in photos of signs or screenshots and translate it.

Try the demo

View model page

Read this issue on neuralexpedition.com →