Typhoon vs OpenThai vs Pathumma: Three Thai LLMs Worth Knowing

By Nokka | September 11, 2026

This article was written by AI (deepseek-v4.1-flash) through Hermes Agent, reviewed and edited by Nokka.

The first question Thai users ask when starting with AI is often "does this need a paid API?" The conversation usually ends with a foreign subscription. The other answer is that Thai teams build open-weight models you can download free and use commercially.

Three projects matter today. Each is strong in a different place, and picking wrong means wasted setup time.

Typhoon: the widest research portfolio

Typhoon is a project from SCB 10X, positioning itself as a frontier AI research lab in Thailand [1].

The numbers they publish: 30 million API calls, 2.6 million downloads, and 13,000 unique users [1].

What separates Typhoon from the other two is breadth. It is not just a chat model. There are reasoning models, speech recognition models, document reading models, and Thai-English translation models [1].

The most interesting recent work is Typhoon Isan ASR, which recognizes Isan speech. Almost nobody else works on this, because regional dialects are something foreign models skip entirely [1].

OpenThai: the most permissive licensing

OpenThai (formerly OpenThaiGPT) is run by the Thai AI Entrepreneur Association with iApp Technology and a volunteer community [2].

The clearest differentiator is licensing. The flagship models use Apache 2.0, allowing commercial use without asking permission and without API fees [2].

The current flagship is OpenThai 2.0 at 27 billion parameters. It reads images, reads Thai handwriting, and handles agent-style tool calling better than the base model it was adapted from [2].

There is a specialized variant worth noting: OpenThai 2.0 Legal at 30 billion parameters in a mixture-of-experts layout, activating only 3 billion per token, so it runs on a 24 GB GPU. Its standout is recalling legal text from memory and answering with the law name and section number in a format you can feed into other systems [2].

Pathumma: academic focus on cultural understanding

Pathumma is developed by a Thai research team at VISTEC, positioned as an open model addressing the rigidity of closed models [3].

What differs from the first two is the focus on cultural and Thai contextual understanding, not just grammatical correctness. This is the recurring problem when foreign models read Thai text with social nuance.

Side by side

	Typhoon	OpenThai	Pathumma
Owner	SCB 10X	AIEAT + iApp	VISTEC research team
License	Open	Apache 2.0	Open
Main strength	Broad portfolio (speech/docs/translation)	Thai documents + Thai law	Thai cultural understanding
Runs locally	Yes	Yes (Ollama/LM Studio)	Yes
Regional dialects	Yes (Isan)	No	No
Best for	Teams needing many capabilities	Legal and document work	Language and culture research

Before you choose

One All three are far smaller than frontier foreign models. Against GPT-6 or Claude Opus, multi-step reasoning is still behind. Where they win is Thai-specific work and cost.

Two "Open" does not mean the same thing across projects. Read the license before commercial use. OpenThai states Apache 2.0 clearly, but check the others carefully.

Three Running locally needs adequate hardware. Models at 27 billion parameters and up require substantial GPU or RAM. On a small machine, start with a 7B variant.

Four All benchmark numbers come from the projects themselves. There is no independent evaluation comparing all three on the same test set yet. Try them on your own actual work before deciding.

From someone writing Thai professionally with AI

I write Thai content with AI daily, and I keep hitting a problem foreign models handle poorly: phrasing translated straight from English. Grammatically correct sentences nobody actually says.

What especially interests me is Typhoon Isan ASR, because frontier labs skip regional speech work for business reasons. Little data, poor return. A Thai team doing it anyway is worth more than any download count.

The honest counterweight is development speed. Thai projects have much smaller teams, so update cycles are longer. Competing with vendors shipping new models monthly is structurally unfavorable.

My advice: use Thai models for what they genuinely win, Thai documents, Thai law, Thai speech. For maximum capability work, foreign models remain the tool. That is not a defeat. Knowing where you win is basic to working well.

If you want to start without installing anything, both OpenThai and Typhoon have free web playgrounds [1][2].

References

[1] SCB 10X, "Typhoon — Thailand's Frontier AI Research Lab" (accessed Sep 11, 2026), https://opentyphoon.ai/

[2] AIEAT & iApp Technology, "OpenThai — Open Thai Language Models" (accessed Sep 11, 2026), https://openthai.aieat.or.th/

[3] Thai NLP Community, "Thai large language models benchmarks for cultural and core capabilities" (2026), https://www.facebook.com/groups/thainlp/posts/2180375852343893/