Skip to main content

OpenThai 2.1 OCR: phone photos, formulas, and fixes from the people who use it

· 7 min read
Kobkrit Viriyayudhakorn
CEO @ iApp Technology

Six days ago we released OpenThai 2.0 OCR, two free Thai OCR models from OpenThai (formerly OpenThaiGPT). People put them to work at once: a web page that reads documents in the browser, desktop apps for Windows, and tests on close to a thousand pages of Thai government documents. They also told us, in comments and in their own code, where the models went wrong. Today we release version 2.1 of both models, and most of what changed comes from those reports.

OCR Markdown: huggingface.co/iapp/OpenThai2.1-OCR-Markdown-20261118 · OCR Text: huggingface.co/iapp/OpenThai2.1-OCR-Text-20261118 · GGUF for llama.cpp: OpenThai2.1-OCR-Markdown-20261118-GGUF

OpenThai 2.1 OCR Markdown​

OCR Markdown reads a whole page into Markdown, with every table kept as an HTML table. Version 2.0 was trained mostly on Thai official documents. Version 2.1 adds photographed pages, bent, creased and shadowed, and nine more page types: academic papers, books, comics, exams, magazines, newspapers, reports, slides and textbooks, in Thai and English.

Public benchmarks of phone photos and English pages, run with Typhoon OCR V2 2B's public evaluation recipes and scored with each benchmark's own scorer. The best score in each column is in bold.

ModelSizeMDPBench, Thai and EnglishMDPBench, ThaiolmOCR-benchOmniDocBench v1.6
OpenThai 2.1 OCR Markdown0.63B82.580.773.179.6
OpenThai 2.0 OCR Markdown0.63B73.772.750.961.8
Nemotron 3.5 Parse (base model)0.63B79.677.462.480.1
Typhoon OCR V2 2B2B87.085.084.791.5
  • MDPBench: 320 Thai and English pages, three quarters of them phone photos, without the formula metric. −4.5 against Typhoon OCR V2 2B, [−7.0, −2.0], and −4.3 on Thai pages, [−8.4, −0.4]. The gap is largest on photographed Thai pages (78.4 against 84.4) and their tables; born-digital Thai pages are level (87.7 against 87.1).
  • olmOCR-bench (1,403 English PDF pages) and OmniDocBench v1.6 (1,651 English and Chinese pages): most of the gap to Typhoon OCR V2 2B is formulas, where it scores 67.0 against 95.1 (OmniDocBench formula CDM). Chinese is not a target language.

On phone photos (MDPBench) the score rises from 73.7 to 82.5, and on Thai pages alone from 72.7 to 80.7. On English PDF pages (olmOCR-bench) it rises from 50.9 to 73.1.

Thai official documents read as they did. On 65 Thai government and university pages with tables it scores 94.0 (2.0: 95.2, [−2.9, +0.3], not a significant difference), still ahead of Typhoon OCR V2 2B at 91.6. The model has the same size, 0.63 billion parameters, and the same speed, 1.5 seconds a page on one NVIDIA H100.

The Markdown itself is cleaner, again from what users reported:

  • Formulas come out as LaTeX between $$.
  • Page headers and footers stay out of the Markdown. They are still in the list of blocks, with their boxes.
  • Blank pages come back empty. Version 2.0 sometimes invented text on an empty sheet.
  • ำ is written as one character, so Thai text matches what search and spell-checking expect, and table cells no longer carry rowspan="1" or colspan="1".

Now on a Mac or a CPU, with llama.cpp​

Version 2.0 needed vLLM, which runs on Linux. Without it, people reported 5 to 11 seconds a page on a laptop GPU and one to two minutes on a CPU. Version 2.1 also comes as GGUF files for llama.cpp, so it runs on a Mac, on a CPU or on a consumer GPU without PyTorch or vLLM. The model is not in upstream llama.cpp yet, so we publish our fork: iapp-technology/llama.cpp, branch nemotron-parse.

The GGUF version's Markdown was identical to the original model's on 30 of 30 test pages. On an Apple M3 Max a page takes 4.8–7.0 seconds with Metal, and about 58 seconds on the CPU alone. The commands are on the GGUF repository and in our guide for Windows, Mac and Linux.

The API runs 2.1​

The OpenThai 2.1 OCR Markdown API has run version 2.1 since 08:33 on 18 November 2026, with the same endpoint and the same price, 0.15 IC a page. If you use it, four things in the Markdown change: page headers and footers are no longer in it (they stay in blocks), formulas come as $$...$$, blank pages come back empty, and ำ is one character.

OpenThai 2.1 OCR Text​

OCR Text is the small model: three files, 6.6 MB in all, that read a page in 0.39 seconds on a CPU, without a GPU or an internet connection. Version 2.1.4 keeps 2.0's model files, with a new runtime and a second recognizer. The fixes:

  • Old books whose tone marks float above the line. On a test page from a book printed in 1935, the error rate fell from 21.5% to 4.8%.
  • Loopless fonts such as Kanit, Prompt and Sukhumvit Set no longer come out as Latin letters: on 12 test pages in fonts the model never saw, stray Latin letters fell from 58 to 11.
  • Two- and three-column pages are read column by column, and lists keep their row order.
  • One-character table cells: a quantity "2" or a day code M, T or W alone in a cell is kept, and amounts such as "280,00" come out as "280.00".
  • Less memory on Mac and Windows, and a requirements.txt with the tested package versions.

On the 12 printed benchmark sets the average moves from 87.8 to 88.3, and on ThaiOCRBench full pages from 82.0 to 84.1. Nearly every fix above started as a report: a comment on our Facebook post, or a workaround in a project on GitHub built on the model.

The OpenThai 2.1 OCR Text API has run version 2.1.2 since 16 November 2026: the tone-mark, column and loopless-font fixes. The one-character cells and decimal points of 2.1.4 are in the download.

Where it is still behind​

On phone photos, English pages and formulas, Typhoon OCR V2 2B is still ahead of OCR Markdown: 87.0 on MDPBench against our 82.5, 84.7 on olmOCR-bench against 73.1, and 91.5 on OmniDocBench against 79.6. The widest gap is formulas: 67.0 against 95.1 on OmniDocBench's formula score. For the next version we are collecting real photographs of Thai documents and more formula data.

Upgrading from 2.0​

Both repositories now carry the release date in their names. The old names, iapp/OpenThai2.0-OCR-Markdown and iapp/OpenThai2.0-OCR-Text, redirect to them, so existing links and code keep working and get version 2.1.

hf download iapp/OpenThai2.1-OCR-Markdown-20261118 --local-dir OpenThai2.1-OCR-Markdown
hf download iapp/OpenThai2.1-OCR-Text-20261118 --local-dir OpenThai2.1-OCR-Text

To stay on a version, pin it: --revision v2.1.0 for OCR Markdown 2.1 or --revision v2.1.4 for OCR Text 2.1.4, and --revision v2.0.0 for either model's 2.0. In Python, pass the same value as revision= to snapshot_download. To keep 2.0's Markdown from the 2.1 model (page headers kept, formulas as \(...\)), see the model page.

Thank you to everyone who tried 2.0 and told us what went wrong. If a document still reads badly, write to sale@iapp.co.th with a sample that holds no personal data.

Built on NVIDIA Nemotron. OCR Markdown is a Thai fine-tune of NVIDIA Nemotron 3.5 Parse under the OpenMDW-1.1 licence; OCR Text is Apache-2.0. Developed by iApp, launched with AIEAT. Trained on NVIDIA H100 GPUs provided by Siam AI Corporation.