Measured
Compared with a leading cloud OCR service (Cloud OCR A) on 4 October 2026: the same page images and the same checked answers for both, scored by character error rate (lower is better). None of these documents was used to train the model.
| Thai documents | Cloud OCR A | iApp | Better on: iApp / Cloud OCR A |
|---|
| Budget documents, parliament (539 pages) | 37.4% | 17.2% | 469 / 25 pages |
| Royal Gazette (221 pages) | 13.8% | 6.1% | 217 / 0 pages |
| SEA-DocBench, Thai (1,225 pages) | 49.4% | 48.7% | 443 / 243 pages |
| Old books, Wikisource (608 pages) | 25.4% | 21.0% | 413 / 124 pages |
| Administrative Court judgments, scanned (150 documents) | 6.2% | 3.1% | 150 / 0 documents |
| All 2,825 benchmark pages | 36.0% | 29.6% | 1,654 / 425 pages |
The court judgments are within the model's training domain: it was trained on other judgments of the same court, never on these. The gap is widest on government documents: the cloud service reads a two-column budget table as all the labels and then all the amounts, so each amount loses its line item, while iApp reads it row by row and its text can be used as it is.
The court benchmark of the August whitepaper (20 judgments, 242 pages, the same script and scoring), re-run on the production service on 5 October 2026:
| Measure | v3.0.0, Oct 2026 | v2.1.13, Aug 2026 | Cloud OCR A |
|---|
| Character error rate | 3.7% (96.3% correct) | 5.2% (94.8%) | 6.1% (93.9%) |
| Word error rate | 5.6% (94.4% correct) | 7.0% (93.0%) | 7.5% (92.5%) |
| Layout analysis | 96.97% AP, the same layout model | 96.97% AP | |
Measured on the production service in October 2026: a median of 0.60 s per page of plain text; about 7.5 pages a second (27,000 an hour) for plain text, and about 0.6 pages a second (2,100 an hour) for layout, Word, PowerPoint and Excel. The service runs on two servers and stays up if one fails.
Method, every set with its confidence interval, and the per-document results: the benchmark paper (PDF)
Limits
- Printed text only; handwriting is not read.
- Scans at 200 dpi or higher give the best results.
- Short photographed text, such as signs and scene text, is read, but the cloud service was slightly more accurate on it in our tests.
Data handling
Uploaded documents are processed on iApp servers in Thailand and used only to produce the requested output; generated links expire after 10 minutes and generated files are deleted within one hour. Consent and notice under PDPA and GDPR remain with you as the data controller. An on-premise deployment, in which no document leaves your network, is available: contact us.
Changelog
| Version | Date | Changes |
|---|
| v3.0.0-20261005 | 12 Nov 2026 | Renamed from General Thai Document OCR to OpenThai 2.0 OCR Text, after the model it is built on. The service, endpoints, requests, responses and prices are unchanged. |
| v3.0.0-20261005 | 5 Oct 2026 | iApp's own Thai OCR model reads the pages; on Thai documents it is more accurate than a leading cloud OCR service (measured), and the character error rate on the August benchmark falls from 5.2% to 3.7%. The service runs on two servers: about 27,000 pages an hour, no downtime on deploys. Plain text from 1 IC to 0.049 IC per page; layout, Word, PowerPoint, Excel and searchable PDF from 1 IC to 0.15 IC per page. Endpoints, requests and responses are unchanged. |
| v2.1.13-20260822 | 22 Aug 2026 | New engine (character error rate down 15% relative, median page time from 3.33 s to 0.53 s), signed expiring file links, on-premise deployment, and PowerPoint, Excel and searchable PDF out of preview. Benchmark paper (PDF) |
| v2.1.12-20251116 | 16 Nov 2025 | Added the iapp-input-pages response header. |
| v2.0 | Aug 2022 | Document layout analysis and Word output. |
