Skip to main content

Open weights

OpenThai OCR

Two open models that read Thai and English documents on your own computer: one returns the text of printed pages on any CPU, the other reads whole pages with their tables.

  • Models2
  • Text, one CPU worker0.39s a page
  • Thai document tables94.0score
  • DownloadFree

OpenThai 2.1 OCR TextOpenThai 2.1 OCR Markdown

Version 2.1, released 18 November 2026 by iApp Technology as open weights on Hugging Face. Version 2.0 (12 November 2026) stays available under the tag v2.0.0.

Which one to use​

Use OCR Text for the text of printed pages, on any computer and at volume. Use OCR Markdown for whole pages whose tables must stay tables, for phone photos, and for handwriting.

OpenThai 2.1 OCR TextOpenThai 2.1 OCR Markdown
ReturnsThe page's text in reading order, with a box and a confidence for every lineMarkdown, with tables as HTML (merged cells kept) and formulas as LaTeX
Size1.60 million parameters, 6.6 MB0.63 billion parameters
Runs onAny CPU: 0.39 s a page, no GPUOne GPU: 1.5 s a page (median, H100); also a Mac or a CPU
Printed Thai, average of 12 sets88.390.2
Thai document tables, 65 pagesNot read as tables94.0
LicenceApache-2.0OpenMDW-1.1
WeightsHugging FaceHugging Face

Scores are character accuracy, 100 × (1 − character error rate), and table similarity; the sets, the method and the other engines tested are on each model page.

Run them​

Releases​

From OpenThai (formerly OpenThaiGPT). Developed by iApp, launched with AIEAT. Trained on NVIDIA H100 GPUs provided by Siam AI Corporation.