Before AI can help someone work with a scanned document, it needs to read it correctly. For teams handling Thai laws, safety procedures and handwritten records, that first step matters. A missed word or misread number can change what a document means.
With schols-ocr-bench, we are sharing our work to make that choice more informed. Our public research project compares how well different systems read and understand Thai documents, giving teams evidence they can inspect rather than relying on a product demonstration.
The project builds on ThaiOCRBench, the research benchmark by Nonesung and colleagues. Their work brings Thai text and document understanding into focus, including challenges with handwriting, document structure and invented content.
Our contribution, ThaiOCRBench Small, selects 240 samples across eight tasks, with 30 examples per task. Each run uses the same selection, making repeat comparisons easier. We tested nine systems from Google, OpenAI, Anthropic, iApp and Typhoon, covering printed text, full pages, handwriting and questions about visual information. See the project and methodology.
In our February 25, 2026 evaluation, Gemini 3 Pro recorded the highest average score at 0.815, followed by Gemini 3 Flash at 0.813. Flash led on full-page text extraction and handwriting, scoring 0.978 and 0.856 respectively. These are benchmark scores, not percentages of documents read perfectly. Explore the results.
The smaller test makes repeated evaluation more accessible. We publish the code and saved results, and the runner can resume completed work. The code is available under the MIT licence, so other teams can inspect and adapt it.
The results are a dated snapshot covering eight of the original thirteen tasks. Differences in scoring, image-size limits affecting Claude and connection errors affecting Typhoon also matter when interpreting them. Read the comparison limits.
For Schols, this work reflects a practical priority: better evidence for the tools used to read Thai documents. We invite researchers and teams building Thai-language software to explore the project, repeat the tests and build on the work.