Pathumma LLM
Thailand's government-built Thai model family for text, documents and speech
Overview
Pathumma LLM is a family of open Thai-language models from NECTEC, part of Thailand's national science agency NSTDA, covering text, vision, OCR, speech and audio. It is offered as free downloadable weights on Hugging Face, an API and hosted apps such as DocChat, DocGen and Partii Note through the AI for Thai portal. Recent text models are post-trained from ThaiLLM, Thailand's national open-weight base 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 These figures reflect RECATOOLS research as of 19 Oct 2026 and may be out of date or incomplete.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 Pathumma LLM
- Download open Thai text, vision, OCR, speech-recognition and text-to-speech models from Hugging Face
- Question long Thai documents (text, .docx, PDF or a website) through DocChat
- Draft official Thai letters, reports and terms of reference with DocGen
- Transcribe and summarise Thai meetings and lectures with Partii Note
- Call the models from Python through the AI for Thai API and its aift package
- Route Thai and English prompts by data sensitivity with the cross-border guardrail model
ASEAN Perspective
Pathumma LLM in Southeast Asia
Pathumma is a multimodal model from NECTEC that processes Thai speech as well as text and images. This matters in a country where much official business is conducted orally and where general-purpose models struggle with Thai audio. Unlike a pilot project, it's in active use: the Thai Parliament uses it to summarise meetings and legislative drafts. It's a rare public example of a government research institute successfully shipping a multimodal model tuned to a national language.
Pathumma is the most complete public-sector Thai AI stack available: open text, vision, OCR and speech models plus ready-made tools for Thai documents and meeting transcripts, kept active through 2026 with OCR, reasoning and guardrail releases. For a Thai agency or developer who needs to keep data in the country, it is a sensible place to start.
Check each model's licence before shipping, since they differ and two declare none. Expect to integrate several checkpoints yourself, and benchmark against Typhoon or OpenThaiGPT on your own tasks, because nearly all published results are NECTEC's.
What we found
Pathumma LLM (ปทุมมา) is the Thai-language AI programme of NECTEC, the computing research centre inside Thailand's national science agency, NSTDA. It is a family of separately released models rather than one model: text, vision and audio models first published on Hugging Face in October 2024, since joined by Thai speech recognition, text-to-speech, OCR models and a data-classification guardrail. NECTEC's Hugging Face organisation lists 20 models. Most are Apache 2.0, but not all: two older vision releases declare no licence, a research model inherits Gemma's terms, and the newest speech recogniser is CC BY 4.0.
The foundations have shifted. Version 1.0 of the text model was a 7-billion-parameter fine-tune with Qwen2.5 lineage. Later 8B releases are post-trains of ThaiLLM, a national open-weight base model built by NECTEC with BDI, VISTEC, AIEAT and AIAT; SCB 10X's Typhoon-S, OpenThaiGPT and KBTG's THaLLE post-train the same base, so these Thai models are siblings as much as rivals. The August 2026 text release, version 4.0.0, is a 4B reasoning model whose card reports 85.0 on a Thai MATH-500 and 56.67 on a Thai AIME 2024, NECTEC's own figures.
Government use is the point. NECTEC's June 2025 page lists DocChat for questioning documents, DocGen for drafting official Thai letters and terms of reference, and Partii Note for transcribing and summarising Thai speech, all reached through the AI for Thai portal alongside an API. A March 2025 NSTDA statement said the Thai Parliament had adopted Pathumma to summarise meetings and legislative drafts; nothing found since confirms it. In August 2026 Post Today reported plans to make Pathumma a central agent linking government chatbots.
Caveats are real. The 2024 audio model calls itself experimental and accepts 30 seconds of audio, the hosted text demo warns that general chat accuracy is limited, and almost every published result comes from NECTEC. The latest release, on 23 September 2026, is a small guardrail that sorts Thai and English text into seven sensitivity classes for routing under Thailand's data protection law, reporting 92.8% accuracy on a synthetic test set.
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 Pathumma LLM's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Pathumma LLM 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 Pathumma LLM directly →
Spotted something out of date? Suggest an update →
Alternatives to Pathumma LLM
More in LLMs & Chat
