OpenThai-SystemOne

Typed decisions on Thai and English text: send a state and questions, and get a calibrated probability for every option in one forward pass, with no text generated.

POST/v3/store/openthai/systemoneTry it

• PriceFree
• Served20.9Kcalls

• v0.3-20260922
• Active

Input

state

ticket: โดนหักเงินซ้ำสองครั้งเมื่อวานนี้ ขอเงินคืนด่วนนะครับ โทรไปสามรอบแล้วไม่มีใครรับ

questions

department: type: choice instructions: ทีมใดควรรับผิดชอบ criteria: { billing, technical, sales } frustration: type: score instructions: ลูกค้าหงุดหงิดแค่ไหน criteria: ใจเย็น หงุดหงิดแต่สุภาพ โกรธมาก refund: type: noul instructions: ลูกค้าขอเงินคืนอย่างชัดเจนหรือไม่

OutputPOST /v3/store/openthai/systemone

model

openthai-systemone

answers

department: type: choice choice: billing probabilities: { billing, technical, sales } confidence: 0.939 abstain: 0.003 frustration: type: score score: 1.974 legend: { 0, 1, 2 } probabilities: { 0, 1, 2 } confidence: 0.895 refund: type: noul noul: 0.989

usage

input_tokens: 175 output_tokens: 0 permutations: 1

Measured on 4 October 2026, probabilities rounded: one forward pass, 175 input tokens, 0 output tokens.

Every field, under Response

Request

curl -X POST 'https://api.iapp.co.th/v3/store/openthai/systemone' \
  -H 'apikey: YOUR_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"state": {"ticket": "โดนหักเงินซ้ำสองครั้งเมื่อวานนี้ ขอเงินคืนด่วนนะครับ โทรไปสามรอบแล้วไม่มีใครรับ"}, "questions": {"department": {"type": "choice", "instructions": "ทีมใดควรรับผิดชอบ", "criteria": {"billing": "การเงิน/คืนเงิน", "technical": "ระบบใช้งานไม่ได้", "sales": "สอบถามสินค้าและราคา"}}, "frustration": {"type": "score", "instructions": "ลูกค้าหงุดหงิดแค่ไหน", "criteria": ["ใจเย็น", "หงุดหงิดแต่สุภาพ", "โกรธมาก"]}, "refund": {"type": "noul", "instructions": "ลูกค้าขอเงินคืนอย่างชัดเจนหรือไม่"}}}'

Response

200OK. The fields are explained under Response.

{
  "model": "openthai-systemone",
  "answers": {
    "department": {"type": "choice", "choice": "billing",
      "probabilities": {"billing": 0.989, "technical": 0.005, "sales": 0.006},
      "confidence": 0.939, "abstain": 0.003},
    "frustration": {"type": "score", "score": 1.974,
      "legend": {"0": "ใจเย็น", "1": "หงุดหงิดแต่สุภาพ", "2": "โกรธมาก"},
      "probabilities": {"0": 0.004, "1": 0.018, "2": 0.978},
      "confidence": 0.895},
    "refund": {"type": "noul", "noul": 0.989}
  },
  "usage": {"input_tokens": 175, "output_tokens": 0, "permutations": 1}
}

Request

JSON. state is what the decision is about, and questions maps names of your choosing to typed questions, all answered in the same forward pass. A request carries at most 64K tokens.

Headers

apikeystringrequired

Your API key, created under API Keys in the control panel.

Body

statestring or objectrequired

What to decide about: text, or any JSON, such as a support ticket, a comment, a document or a UI accessibility tree.

questionsobjectrequired

A map of question name to question.

questions.typestringrequired

choice to pick one of named options, score to place on an ordered scale, noul for yes or no.

questions.instructionsstringrequired

What to decide, in Thai or English.

questions.criteriaobject or array

choice: up to 255 option names, each with a short description. score: 2 to 10 level labels, in order. Not used by noul.

order_invariantboolean

Averages each choice answer over several orders of its options, which removes sensitivity to option order; about twice the latency. On by default for choice questions with more than 10 options.

permutationsinteger

How many option orders to average when order_invariant is on.

Price

Free during the preview: 0 IC per request with any iApp API key, up to 100 requests a minute and 1,000 decisions a day per key. Per-decision pricing will be announced; the free preview continues.

Response

JSON with one answer per question, under the name you gave it, as in the panel. The model generates no text, so usage.output_tokens is always 0. A malformed request, such as a score with fewer than 2 levels or a choice with more than 255 options, returns 422.

answersobject

One object per question.

answers.typestring

The question's type.

answers.choicestring

For choice: the most probable option.

answers.scorefloat

For score: the probability-weighted level index, which can be fractional; legend maps each index to its label.

answers.noulfloat

For noul: the probability of yes.

answers.probabilitiesobject

The probability of each option or level; they sum to 1.

answers.confidencefloat

1 minus the normalised entropy. Send low-confidence answers to a larger model or a person.

answers.abstainfloat

For choice: the probability that none of the offered options fits.

usageobject

input_tokens, output_tokens (always 0), and permutations, the number of option orders averaged.

Details

Measured

Measure	Result
Bespoke Labs' public System One benchmark, 13 subsets, macro average	74.3, against 74.8 for Bespoke-Nimble-9B and 76.0 for Jev 1.13.0
Thai intents, MASSIVE-th (60-way), held out from training	90.0% accuracy, calibration error 0.043
Thai news topics, Prachathai67k, held out from training	98.1% accuracy, calibration error 0.004
Round trip through the gateway, three-question Thai ticket	0.23 s
One H100 GPU, batch 1, three questions	40 to 70 ms
Output tokens	0

Per-subset results, the other Thai test sets and the calibration tables are on the model page and in the model card.

Limits

• 100 requests a minute and 1,000 decisions a day per key during the free preview; 255 options per choice, 2 to 10 levels per score, 64K tokens per request.
• Version 0.3, a 0.8B model, text only. It answers only the options you offer and can still choose wrongly; it is weakest on 5-level rating tasks, PubMedQA and Thai social-media sentiment.

Open weights

The same API runs on your own machine from the open weights, and answers POST http://localhost:8000/v1/systemone with the same body:

pip install "git+https://github.com/iapp-technology/openthai-systemone"
OPENTHAI_SYSTEMONE_MODEL=iapp/OpenThai-SystemOne uvicorn openthai_systemone.server:app --port 8000

Changelog

Version	Date	Changes
v0.3-20260922	22 Sep 2026	5,000 more fine-tuning steps on real train splits and targeted synthetic sets, re-calibrated: public macro average from 63.2 to 74.3. The hosted API serves v0.3.
v0.2-20260921	21 Sep 2026	Thai social-media sentiment training, re-calibrated; order_invariant and permutations added. Public macro average from 61.9 to 63.2.
v0.1-20260920	20 Sep 2026	Launch: the hosted API as a free preview, and the open weights and training recipe under Apache 2.0.