Typhoon, OpenThaiGPT, and Pathumma are Thai-focused model families, not three directly interchangeable versions of one product. Typhoon comes from SCB 10X and has evolved across text and multimodal releases; OpenThaiGPT 1.5 offers several Qwen 2.5-based chat sizes with retrieval and tool features; Pathumma, from NECTEC under NSTDA, is designed as a Thai-first multimodal model for text, images, and audio. There is no shared, current benchmark here that establishes an overall winner, so choose by the exact release, task, modality, and deployment route.
How the three Thai LLM families differ
| Family | Developer and release context | Documented strengths | What to verify |
|---|---|---|---|
| Typhoon | Developed by SCB 10X. The January 2024 launch covered a Thai-optimized 7B model in pretrained and instruction-tuned variants. SCB 10X’s January 2025 overview describes later 1.5/1.5X releases and the Typhoon 2 family, including multimodal development. | Thai-language text models across a changing family; later releases broaden the family beyond the initial 7B research release. | Check the exact model card for version, modality, license, and serving requirements. Capabilities from one Typhoon release should not be assumed for another. |
| OpenThaiGPT / OpenThai | The OpenThai project presents OpenThaiGPT 1.5 in 7B, 14B, and 72B variants based on Qwen 2.5. | Official 1.5 materials list multi-turn Thai chat, retrieval-augmented generation (RAG), tool calling, and a stated long context window. | Use “OpenThaiGPT 1.5” when relying on those details. Check the specific release for current naming, context limit, license, and access route. |
| Pathumma | Developed by NECTEC, part of Thailand’s National Science and Technology Development Agency (NSTDA). | The official project describes a Thai-first multimodal model for text, images, and audio, with tasks including question answering, summarisation, drafting, translation, and RAG over Thai knowledge. | Confirm which specific release supports the modality and task you need, along with its license and deployment route. |
These are families whose releases differ in size and capability. The table summarizes documented descriptions, not hands-on testing or a claim that every listed feature is present in every version.
As an Amazon Associate I earn from qualifying purchases.
Which one fits your task?
For Thai text chat, writing, or question answering
Start with the exact text-oriented release you can access, then test it on representative Thai prompts from your own workload. Compare answer accuracy, instruction following, fluency, and how it handles domain-specific vocabulary. A general model description or an exam score cannot tell you which release will work best for your particular documents or users.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →For retrieval and tool-connected applications
OpenThaiGPT 1.5’s official materials explicitly list RAG and tool calling. Treat these as model features, not a complete application: a working retrieval system still needs an indexed knowledge source and application logic, while tool use requires you to connect and control the relevant tools. Check the exact version’s documentation before designing around either feature. Pathumma’s official site also lists RAG among its Thai-language use cases.
For image or audio input
Pathumma’s official description explicitly covers text, images, and audio. SCB 10X’s Typhoon 2 overview discusses multimodal development, but that does not establish that every Typhoon release handles every modality. Select a named release and confirm the precise input and output modes it supports before committing to an image- or audio-based workflow.
For a particular size or deployment environment
OpenThaiGPT 1.5 is described in 7B, 14B, and 72B sizes, giving teams several model scales to evaluate. That alone does not establish which will fit a given server: actual memory and compute needs depend on the specific release, quantisation, runtime, and serving setup. The sources do not validate a common hardware configuration across all three families, so confirm requirements from each model card and test on the intended stack.
What the benchmark claims do—and do not—show
Typhoon’s 2.62-times figure
SCB 10X’s January 2024 announcement said the original Typhoon 7B could process Thai text 2.62 times faster than GPT-3.5. The accompanying Typhoon paper describes a 2.62-times efficiency claim for tokenizing Thai text. This is a scoped claim about the original release’s Thai text processing/tokenization comparison; it is not evidence that every Typhoon model generates answers 2.62 times faster. SCB 10X’s Typhoon announcement and the Typhoon research paper provide the original context.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →ThaiExam and OpenThaiGPT 1.5 evaluations
The Typhoon paper describes ThaiExam as drawing on examinations for Thai high-school students and investment professionals. That offers evidence about performance on that exam-oriented evaluation, not a universal measure of day-to-day writing, retrieval, or multimodal ability.
OpenThaiGPT 1.5’s paper reports fine-tuning with more than 2,000,000 Thai instruction pairs, and its official page includes benchmark tables for named exams and comparison models. Those claims belong to OpenThaiGPT 1.5 and its documented evaluation material; they should not be combined with unrelated scores as though all models were tested under one protocol. See the OpenThaiGPT 1.5 page and the OpenThaiGPT 1.5 paper.
Why the evidence does not name an overall winner
The documented figures concern different versions, tasks, and evaluation contexts. The reviewed Pathumma description establishes modalities and use cases but does not provide a result under the same benchmark protocol as the cited Typhoon and OpenThaiGPT material. That absence is not evidence that Pathumma performs worse; it means the available figures do not support a like-for-like ranking.
Rank #4
A practical selection checklist
- Name the task. Decide whether you need Thai chat, summarisation, translation, exam or domain question answering, coding, retrieval, or document processing.
- Pin down the release. Record the exact model name and version rather than comparing family names alone.
- Confirm modality. Check that the specific release accepts the image or audio inputs your application requires; do not infer support from a family-level overview.
- Test the same workload. Use the same representative prompts, documents, and scoring criteria for each candidate. If relying on published evaluations, match benchmark, prompt, scoring method, version, and date.
- Check integrations. Determine whether the release documents RAG or tool calling and identify the application-side retrieval, indexing, or tool orchestration you must supply.
- Validate deployment. Review the current model card for license, intended use, hosted API terms, and runtime or compute requirements. Family members may have different terms and technical needs.
These comparisons synthesize public project and paper descriptions; they are not results from a head-to-head test. The relevant starting points are SCB 10X’s Typhoon 2 overview, the OpenThaiGPT 1.5 project page, and the Pathumma official site.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
