Thai AI Firm iApp Releases OpenThai 2.0 OCR: Small Fine-Tuned Models Outscore General Parsers on Thai Government Documents
A 0.63B fine-tuned parser and a 4 MB CPU-only OCR engine, both open-weight, were launched on 12 November and are already in commercial production.
iApp Technology, a Thai AI company, released OpenThai 2.0 OCR on 12 November: two open-weight document models built for Thai government paperwork. The Markdown variant is a 0.63-billion-parameter full fine-tune of NVIDIA's Nemotron 3.5 Parse that outputs structured Markdown with HTML tables, bounding boxes and reading order. The company reports 95.2 on a table-heavy test set and 98.7% character accuracy on Royal Gazette pages from November 2026 - figures that are self-reported. The companion model, OpenThai 2.0 OCR Text, is roughly 4 MB, runs on plain CPU via ONNX at about 950 pages per minute on an 8-core server, and is Apache 2.0 licensed.
The published training recipe - 54,596 rows of synthetic Thai documents, born-digital government files, and pages chosen to break repetition loops - documents a repeatable playbook: a small parser fully fine-tuned on local documents can outscore general 2-4B international models in its own language. For markets with complex tabular government documents, such as Indonesia, that is a directly relevant template, and both weights are already downloadable via Hugging Face.