Skip to main content

Research

Measured on Thai data, and published so anyone can check it

Open models trained for Thai, the benchmarks behind our API figures, and our papers.

Read the benchmarksPublications

Open models

Weights anyone can download and run, with the scores we measured.

Thai-English translation that follows your rules, open at 4B, 2B and 0.8BChindaMTSep 29, 2026Open weightsA 0.8B model that decides instead of writing: a calibrated probability for every option in one passOpenThai-SystemOneSep 20, 2026Open weightsOne open model that reads Thai documents and handwriting, knows Thailand, and does agentic workOpenThai 2.0Aug 27, 2026Open weightsA Thai legal model that knows the statutes by heart, and cites the section it usedOpenThai 2.0 LegalJul 24, 2026Open weightsA Thai model small enough to run on your own laptopChindaLLM 4BMay 28, 2025Open weights

Benchmarks

Each whitepaper states its dataset, method and errors, so anyone can run it again.

Thai vehicle make and model recognition: a benchmark on today’s trafficOn 853 held-out photographs of today’s Thai traffic, version 3.0 names the vehicle brand on 99.4% of images and the exact make and model on 99.18%.99.4% brand accuracyAug 28, 2026Whitepaper, PDF
Passport OCR95.5% verified readsAug 25, 2026PDFThai National ID Card OCR98.0% accuracyAug 24, 2026PDFThai Bank Book OCR96.8% account numberAug 23, 2026PDFThai Driver License OCR95.14% accuracyAug 23, 2026PDFPhotographed Thai National ID Card With Signature Detection97.1% accuracyAug 23, 2026PDFGeneral Thai Document OCR94.8% accuracyAug 22, 2026PDF

Publications

Papers from our research team, newest first.

OpenThaiGPT 1.6 and R1: Thai-centric open-source and reasoning large language modelsarXiv2025LLaVAC: fine-tuning LLaVA as a multimodal sentiment classifierarXiv2025OpenThaiGPT 1.5: technical reportNLDB 20242024

All publicationsDatasets