OpenThai
Thailand's community-built open-weight Thai language models, renamed from OpenThaiGPT
Overview
OpenThai, called OpenThaiGPT until July 2026, publishes free open-weight Thai language models, run by the AI Entrepreneur Association of Thailand (AIEAT) with iApp Technology. Its current models are OpenThai 2.0, a 27B image-and-text model; OpenThai 2.0 Legal for Thai law; and the small OpenThai-SystemOne classifier. Older 1.0, 1.5, 1.6 and R1 models remain downloadable, under licences that differ by model.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 19 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with OpenThai
- Self-hosted Thai chat and document Q&A with OpenThai 2.0, including reading Thai scans and handwriting
- Thai legal answers citing statute names and section numbers from OpenThai 2.0 Legal
- One-pass text classification with confidence scores from the 0.8B OpenThai-SystemOne
- Agent tool-calling in Thai from a single open model
- Local runs through Ollama, LM Studio or vLLM, with GGUF, MLX and quantised builds
- Fine-tuning on Apache 2.0 weights where the model card allows it
ASEAN Perspective
OpenThai in Southeast Asia
OpenThaiGPT comes out of the Thai research community rather than a ministry — AIEAT, AIAT and NECTEC — and it is the model most often cited as the open Thai baseline. A key detail for adoption is its architecture: version 1.5 is a fine-tune of Qwen 2.5. An organisation choosing it for sovereignty is therefore using a Chinese base model with Thai tuning, not a model trained from scratch in Thailand. That may be perfectly acceptable, and it is a different proposition from ThaiLLM or Pathumma. The Apache 2.0 licence and the 7B/14B/72B range mean it can be self-hosted at whatever scale the budget allows.
OpenThai is the most active community source of open Thai models: a 27B multimodal flagship under Apache 2.0, a Thai legal model that already powers a working legal chatbot, and a tiny classifier that runs on a laptop. Launch posts state weaknesses plainly, which makes the benchmarks easier to trust, although nearly all of them are the team's own.
Check each model card's licence rather than the project's blanket Apache claim: several generations carry Llama 2, Qwen or NVIDIA terms. Self-hosting needs real GPUs, and no paid managed service is published. A strong first look for Thai-language work.
What we found
OpenThai is the Thai open-model project that ran as OpenThaiGPT until 19 July 2026, when it renamed itself because, by its own account, OpenAI vetoed its trademark application for the word GPT. It is operated by the Artificial Intelligence Entrepreneur Association of Thailand (AIEAT) with iApp Technology, which does the core development and hosts its API, plus volunteer developers. Siam AI Corporation donated eight H100 GPUs, and a public research grant funds the work. The project began in early 2023, co-founded by AIEAT and the AI Association of Thailand.
The current line-up is three models. OpenThai 2.0, released 27 August 2026, is a 27-billion-parameter image-and-text model on the vendor-named Qwen3.8-27B base, Apache 2.0. Its launch post reports 0.842 on OpenThaiEval, a suite of Thai national exams, and is candid about weaknesses: a dedicated OCR model transcribes better, and handwriting still errs on roughly one character in four. OpenThai 2.0 Legal, announced in July, is a 30-billion-parameter mixture-of-experts model built with NVIDIA tooling as part of the ThaiLLM initiative; NVIDIA's September recap says it powers iApp's Thanoy legal chatbot, used by roughly 43,000 people. OpenThai-SystemOne, released 20 September, is a 0.8-billion-parameter classifier that returns choices with confidence scores and no generated text.
Licensing needs care. The project's licence page says every model is Apache 2.0, but the Hugging Face cards say otherwise: 1.0 carries the Llama 2 licence, 1.5, 1.6 and R1 a Qwen licence, and 2.0 Legal NVIDIA's Open Model Agreement. Read each card before building commercially.
The older generations remain downloadable. The 1.5 paper scored the 72B model at 63.89% on the Thai Exam benchmark, level with GPT-4o and below Claude 3.5 Sonnet; an earlier version of this page gave 64.07% and a Llama 3.1 405B comparison the paper does not contain. Nearly all benchmark figures, old and new, are the team's own, so a Thai deployment should be tested on its own documents before relying on them.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including OpenThai's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with OpenThai unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to OpenThai directly →
Spotted something out of date? Suggest an update →
Alternatives to OpenThai
More in LLMs & Chat
