OpenThai 2.0

What it does​

• Reads Thai documentsHandwriting at 0.261 character error rate against the base model's 0.649; books and Royal Gazette pages at 0.126.
• Answers about documentsReads a document and answers questions on it in one model, with no separate OCR stage.
• Calls tools0.820 on the Berkeley Function-Calling Leaderboard, ahead of its base model and Typhoon 2.5.
• Decodes fasterA bundled draft head gives 75.2 against 50.1 tokens/s on one H100 stream, with identical output.

Measured​

All models were measured by iApp on one suite and one serving setup, each with its own recommended prompts and parameters. Test sets were held out from training by an n-gram leakage check, and the handwriting test shares no text with training. Per-category results are in the Hugging Face repository.

Run it​

import base64
from openai import OpenAI

client = OpenAI(base_url="https://api.iapp.co.th/v3/llm/openthai2p0",
                api_key="YOUR_IAPP_API_KEY")

# Text — ask anything in Thai
r = client.chat.completions.create(
    model="openthai2.0",
    messages=[{"role": "user", "content": "อากรแสตมป์กับภาษีมูลค่าเพิ่มต่างกันอย่างไร"}],
)
print(r.choices[0].message.content)

# Vision — read a Thai document
img = base64.b64encode(open("thai_document.jpg", "rb").read()).decode()
r = client.chat.completions.create(
    model="openthai2.0",
    messages=[{"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{img}"}},
        {"type": "text", "text": "อ่านข้อความในเอกสารนี้ทั้งหมด"},
    ]}],
    temperature=0.0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(r.choices[0].message.content)

curl -s https://api.iapp.co.th/v3/llm/openthai2p0/chat/completions \
  -H "Content-Type: application/json" -H "apikey: YOUR_IAPP_API_KEY" \
  -d '{"model":"openthai2.0","messages":[{"role":"user","content":"สวัสดี"}]}'

ollama run openthai/openthai2.0-qwen3.8-27b

vllm serve iapp/openthai2.0-qwen3.8-27b \
  --max-model-len 32768 --gpu-memory-utilization 0.85 \
  --max-num-seqs 128 --reasoning-parser qwen3 --trust-remote-code

Other formats on Hugging Face​

Settings that matter​

Limits​

• Structured extraction is at parity with the base model: 0.568 against 0.570 on judged ThaiOCRBench. It trails specialist models on tables and key fields; test form pipelines on your own documents.
• Answering questions about scene text in photos (MTVQA-TH) is its weakest reading: 0.819 character error rate against the base model's 0.737. Transcribing scene text is not: 0.035 on the 24 scene-text images of OCR-Eval-104.
• A transcription-only model reads clean print and isolated handwriting better: Typhoon-OCR 1.5 scores 0.014 and 0.054, but takes only a fixed OCR prompt and answers no questions.
• Answers are explanatory by default and terse on request about 40% of the time. For machine-read output, ask for a format, such as "ตอบเป็น JSON เท่านั้น".
• On hard handwriting about one character in four is still wrong, and illegible input can produce invented text. Not evaluated on Thai dialects or on vertical or rotated text.

For tax, legal or medical answers, pair it with retrieval over authoritative sources and human review. For law, use OpenThai 2.0 Legal.

Licence​

Apache 2.0. Fine-tuned from Qwen/Qwen3.8-27B with LoRA (rank 64) on about 143,000 verified rows; the adapter alone (7 GB) is in the repository for serving on top of the base model.

Developed by iApp, launched with AIEAT. Trained and evaluated on 8 NVIDIA H100 GPUs provided by Siam AI Corporation.

@misc{openthai2_2026,
  title  = {OpenThai 2.0: An Open Thai Knowledge and Document AI},
  author = {iApp Technology and Artificial Intelligence Entrepreneur Association of Thailand},
  year   = {2026},
  url    = {https://openthai.aieat.or.th}
}