AI & ML interests

Free, open-weight Thai large language models.

Recent Activity

Organization Card
OpenThai

OpenThai

Free, open-weight Thai large language models.

formerly OpenThaiGPT

Website · Leaderboard · Ollama · GitHub · Discord · Papers


Thai is spoken by 70 million people and served by almost no open models. We build them — trained on Thai data, evaluated on Thai benchmarks, and released open-weight so that Thai organisations can self-host and keep Thai data in Thailand.

Developed by iApp, launched with AIEAT (the AI Entrepreneur Association of Thailand), with a community of volunteer researchers. A contributing member of Thailand's national ThaiLLM initiative.

Which model should I use?

Need Model Size
🎙️ Thai speech to text while people are still talking (80–1120 ms behind) OpenThai 2.0 ASR Streaming ← newest 0.64B
📄 Thai document pages to Markdown, tables stay tables OpenThai 2.0 OCR Markdown 0.63B
🔤 Printed Thai text OCR on a CPU, no GPU OpenThai 2.0 OCR Text 4 MB
🚀 Thai chat, Thai documents and handwriting, agents and tool use OpenThai 2.0 27B
⚖️ Thai law, statute citation, legal RAG OpenThai 2.0 Legal 30B MoE (~3B active)
⚡ Typed decisions in one pass: classification, routing, scoring, yes/no, agent action selection (no text generation) OpenThai-SystemOne 0.8B
🧠 Previous generation: reasoning, maths, code OpenThaiGPT R1 32B
💬 Previous generation: general Thai chat OpenThaiGPT 1.6 · 1.5 72B · 7B / 14B

How they compare with other Thai models, task by task and with sources: openthai.ai/leaderboard.

The OpenThai 2.0 models and OpenThai-SystemOne are published under the iapp org, where they are maintained. They are OpenThai project models — those are the canonical copies, not mirrors. SystemOne: code · playground · pip install openthai-systemone

Papers

Get involved

We need help with Thai evaluation data, domain fine-tunes (medical, government, education), and inference tooling. Start in Discord.


OpenThai (formerly OpenThaiGPT) · Developed by iApp, launched with AIEAT · openthai.ai