API · Open weights
ChindaMT
Thai and English translation that follows your rules.
- Price1IC / 400 characters
- One sentence0.4s
- Native speakers prefer it2 to 1
- Open weights4B · 2B · 0.8B
Published at AACL-IJCNLP 2026, Main Conference. Models and code under Apache 2.0.
ประชุมเลื่อนไปวันศุกร์ เพราะหัวหน้าติดงาน
Without rules
The meeting has been rescheduled to Friday because the supervisor is busy with work.
With a ruleUse a casual tone
The meeting's been pushed to Friday since the boss is swamped with work.
What it does
Translates Thai and English in either direction and follows the rules sent with the text: short instructions in plain English, of the five kinds it was trained on. A rule works best when it is specific and can be checked in the output.
- Thai and EnglishBoth directions, one text or a batch of up to 100,000 characters, each line translated on its own.
- Terminology
Translate บาท as THB,Keep product names in English - Tone and register
Use a formal, polite tone suitable for a business email - Length
Keep it short enough for a subtitle,Use exactly one sentence - Format
Return only the translation, no quotes,Keep the numbering of the list - Language
Avoid English contractions,Write numbers as digits
Measured
How often a judge model chose ChindaMT's translation over a rival's when rules were given, corrected for answer length.
ChindaMT chosenA tie, or the other model chosen
ChindaMT-4B, compared with
- Typhoon-Translate-1.5-4B68.4%
- TranslateGemma-4B87.2%
- MiLMMT-46-4B89.5%
- Native speakersThree native Thai speakers chose ChindaMT-4B over Typhoon-Translate-1.5-4B 2 to 1 in a blind test: 64% of plain translations, 69% of translations with rules.
- Without rules90.7 on FLORES-200 (the mean of CometKiwi, GEMBA-DA and GEMBA-MQM), the highest of the 4B to 9B models tested; 87.8 on WMT24++ news, second to a 7B model at 89.4.
- Speed0.36 s for a sentence, 0.71 s for a page and 1.72 s for five pages, through the gateway from Bangkok.
Methods and the full tables are in the paper.
Fully open
- WeightsChindaMT-4B, 2B and 0.8B, under Apache 2.0
- Training dataChindaMT-Grounded, 1.97 million records
- Evaluation suitesChindaMT-CoreEval and ChindaMT-BroadEval
- Curation codeChindaMT-RGDC
Limits
- Thai and English only.
- It follows rules a good reference translation could satisfy. Strict placeholder tokens and a long mandatory glossary are outside what it was trained for.
- Like any machine translation, it can read fluently and still be wrong: have legal, medical and other high-stakes text reviewed. The API marks the parts most likely wrong.
- 100,000 characters and 2,000 characters of rules per request; 5 requests per second per API key.
On your own servers
The open weights run on your own GPU. For an on-premise deployment with your own terminology, write to sale@iapp.co.th.