The best OCR models in 2026: open-source and vision-language models compared on published benchmarks

On OmniDocBench v1.6, the best OCR model in September 2026 is a small open-source model built for documents. Large general-purpose AI models score lower there. TeleOCR scores 96.91, OvisOCR2 96.47 and PaddleOCR-VL-1.6 96.34. Each is released under Apache 2.0 and has about 1.2 billion parameters or fewer. Parameters measure a model's size, and that's small for an AI model. On olmOCR-Bench, Datalab reports 85.8 for its Chandra OCR 2. That's the highest reported score we found for a model you can download. Both benchmarks use mostly research papers, books and reports, so test any model on your own documents before you pick one.

The scores below come from each benchmark's own page or from the model card, the page where each maker describes its model. We checked them on September 29, 2026, and we didn't run the models ourselves.

The best OCR models in 2026, by published score#

The table lists 14 models in four groups. Most are vision-language models (VLMs), which read a page image and write out its text and layout. Size is the parameter count each maker states, except for Chandra OCR 2, whose count comes from Hugging Face. A model's weights are the trained files you download and run, and the license says how you may use them.

OmniDocBench scores come from its v1.6 leaderboard, updated on September 11, 2026. "Not published" means we found no score for that model on that benchmark. Datalab, which makes Chandra OCR 2, and the makers of dots.mocr ran olmOCR-Bench on their own models. Ai2 runs olmOCR-Bench and also makes olmOCR 2.

Which OCR model fits your job#

Start from the job the model has to do. Then check the license and the hardware you have.

What the two benchmarks measure#

OmniDocBench comes from OpenDataLab, the team that also makes MinerU. Its leaderboard now scores models on 1,651 PDF pages. The pages are research papers, textbooks, exam papers, slides, magazines, newspapers, research and financial reports, and handwritten notes. The overall score averages three parts, for text, tables and math formulas. So a third of the score comes from math formulas, which invoices and bank statements rarely contain.

olmOCR-Bench comes from Ai2, which also makes olmOCR. It runs more than 7,000 pass-or-fail tests on 1,400 documents. These include math papers, old scanned textbooks and letters, dictionary pages and pages with tables. Some tests pass only when page headers and footers are left out of the text. That fits olmOCR's goal of clean, readable text. On a bank statement or an invoice, though, the top of the page often holds the account number or the invoice number. Check that a model trained to drop page headers keeps them. Several scores in Ai2's own table were reported by the model makers, and Ai2 marks them.

The authors of Wild-OmniDocBench point out that standard benchmarks use clean scans and digital PDFs. So they printed OmniDocBench v1.5 pages, bent them and photographed them. They also showed pages on a screen and photographed the screen. Then they scored models on the photos. The first PaddleOCR-VL fell from 91.93 to 72.19, and MinerU2.5 fell from 90.67 to 70.91. The drop was smaller for dots.ocr, at about 10 points, and for Qwen3-VL-235B, at about 9. The benchmark's table covers the first PaddleOCR-VL and MinerU2.5, not PaddleOCR-VL-1.6 or MinerU2.5-Pro. On the same photos, TeleOCR's model card reports 87.36 for the newer PaddleOCR-VL-1.6.

Our take. Use the leaderboards to make a short list of three or four models. Then test them on your worst files and choose the one that does best. Those files are phone photos, faxed forms, 40-page statements and handwritten amounts. A clean PDF only shows how a model handles the easy pages. The pages it gets wrong are the ones your team will fix by hand.

The models in detail#

The model profiles below follow the table's groups. Each fact comes from the model card or the code repository, checked on September 29, 2026.

Small document models

Larger document models

Classic OCR engines

Other models on the boards

General vision-language models score below the five small document models in the table on OmniDocBench v1.6. The best of them is Ovis2.6-30B-A3B, at 93.70. The best closed one is Gemini 3 Pro, at 92.91, and GPT-5.2 scores 86.59. Qwen3-VL-235B scores 89.78.

The board also lists DeepSeek-OCR 2, released under Apache 2.0 in January 2026, at 90.25. The first version of Tencent's HunyuanOCR scores 89.95. Its license grants no rights in the EU, the UK or South Korea, so you can't use it there.

Both boards list an older version of Mistral OCR. The current Mistral OCR 4.1 came out in July 2026, and our Mistral OCR comparison covers what it returns and costs.

What a model returns, and what your process needs#

An OCR model turns a page image into text, Markdown or JSON. Take the bank statement below. It's Harbor Street Bakery's July 2026 statement from First Midwest Bank, and it lists 142 transactions. A good model can return every page as text, with the transactions as Markdown tables.

Your process needs more than that. It needs all 142 rows in one table, in order, with no row lost or repeated where one page ends and the next begins. Dates printed as 07-14 need to come out in one format, such as 2026-07-14. The balances need to add up. The opening 42,180.55, plus 88,412.10 in deposits, minus 91,686.44 in withdrawals, must equal the closing 38,906.21. And a person needs to check any value the model isn't sure of.

Some models cover part of that. GLM-OCR can fill a JSON schema you give it, and PaddleOCR can merge a table that runs across pages. Your team writes and maintains the code for the rest. That code covers field mapping, which matches page text to named fields. It also covers the balance checks, a review screen and the link to your other systems.

How to test an OCR model on your own documents#

Test your short list on your own pages before you choose. Our guide to measuring OCR accuracy covers the metrics.

• Start with your worst pagesPhone photos, skewed scans and faxed pages. On Wild-OmniDocBench, photos cut some 2025 models' scores by about 20 points.
• Include your longest documentsA 40-page statement tests reading order and tables that run across pages.
• Add handwritingFilled-in forms and handwritten amounts show whether a model reads the writing or invents text where it can't.
• Score the fields you useCount the fields a person had to fix, not the characters the model got right.
• Read the license for your useCheck revenue caps, commercial limits and the countries the license covers.
• Time it on your own hardwarePublished speeds come from other GPUs, batch sizes and settings.

If your pages are invoices, not research papers#

The benchmarks are a fair guide for pages like the ones they test, such as research papers and textbooks. For those pages, start at the top of the table and pick the license that fits. For plain printed text on CPUs, PaddleOCR or Tesseract is enough. Invoices and bank statements need more than OCR. They need named fields and a review step, and Docsumo provides both as an intelligent document processing platform, in the cloud.

Book a demo and bring your hardest pages, or start a free trial.

Frequently asked questions#