Photograph a house registration page
Point the camera at a ทะเบียนบ้าน and get every field back as structured data — names, ID numbers, addresses, the lot — without one of them being retyped by a clerk at 4pm.
Multimodal · roadmap Roadmap — not yet shipped
A Thai model that can be shown things.
Most of the Thai language that matters is not typed. It is stamped, photographed, handwritten, printed and scanned again, spoken in a dialect, or read aloud to somebody who cannot read it. A text-only model cannot reach any of it. This is what changes, why Thai specifically needs it, and roughly when.
What a multimodal model is
A text model reads tokens. A multimodal model accepts more than one kind of input — images, audio, video, documents — and reasons across them in the same context as the text. The usual construction bolts an encoder for each modality onto a language model and trains the join, so that a picture becomes something the language model can attend to alongside the words in the prompt.
The term that matters in practice is the vision-language model: a language model with an image encoder, which is what almost every “can it read my screenshot” capability actually is. IBM's definition is the short one — a model that can process and reason across multiple modalities such as text, images and audio — and NVIDIA's glossary is the short one for the vision half.
The idea is older than the current wave. CLIP showed in 2021 that image and text could be trained into a shared space from natural-language supervision; Flamingo showed a frozen language model could be taught to attend to images with a few examples; visual instruction tuning made the result something you could talk to. Everything in production today is a descendant of those three.
Read the primary sourcesWhy Thai needs this more than English does
This is not the same feature in a different language. Four things make Thai document vision a distinct engineering problem, and all four are reasons a general-purpose model underperforms on it.
1
Thai OCR is not solved
No spaces between words, so a recogniser cannot use whitespace to find boundaries. Vowels and tone marks stack up to four levels around one consonant, and a scan that loses the top row silently changes the word rather than corrupting it visibly.
2
The paperwork is photographed, not scanned
The real input is a phone photo of a stamped page on a desk, at an angle, in an office with one fluorescent tube. Clean-scan benchmarks do not predict anything about it.
3
Handwriting is still the interface
Village registers, clinic notes, delivery dockets, school marksheets. Thai handwriting joins and abbreviates in ways that printed-text training does not cover.
4
The numbers are not the numbers
Thai numerals ๐–๙ next to Arabic ones, Buddhist Era years 543 ahead of the Gregorian, and baht amounts written in a Thai numeric register. Read the year wrong and every date in the document is wrong.
Four senses, three models
Shipping as separate models rather than one large one, because the router can then send a photograph to the model that reads and a recording to the model that hears, and neither pays for the other's weights.
image · video frame
ThaiLLM-Vision
Photographs, screenshots, charts, shelves, scenes, forms. Answers about what is in the frame, in Thai, with the region it was served from.
document · handwriting
ThaiLLM-Scribe
The document specialist: stamped, scanned, skewed, multi-column, part-handwritten Thai paperwork into structured fields you can put in a database.
audio in · audio out
ThaiLLM-Voice
Thai speech recognition and synthesis across the four regional varieties, including the common case of a question asked in Isan that has to be answered in Central Thai — or the reverse.
all of the above, together
Cross-modal
The one that matters: a photograph and a spoken question in the same request, answered against your own documents. This is the capability the separate models exist to make cheap.
Eighteen things you could not do before
Every one of these is a job somebody in Thailand does today by typing out what is in front of them. Illustrative — these describe intended capability, not shipped features.
Government and public administration
Read a Royal Gazette notice from a photo
A scanned ราชกิจจานุเบกษา page becomes a summary where each obligation is traceable to the article it came from, and the effective date is read as a Buddhist Era year rather than guessed.
Turn handwritten village minutes into a record
Photograph the page a committee secretary actually wrote, and get typed Thai minutes in official register, with the names spelled consistently and the resolutions numbered.
Check a form before you queue for it
Photograph a half-filled ministry form and be told which boxes are missing, which stamp is absent, and which attachment the office will send you home for.
Health
Explain a prescription label out loud
A photo of a Thai pharmacy label becomes a plain-Thai spoken explanation of the dose and the timing — for a patient who reads slowly, or does not read the register the label is written in.
Make a discharge sheet actionable
Photograph the hospital discharge summary and get “what to do this week” in the dialect spoken at home, with the follow-up date converted from พ.ศ. and put in a reminder.
Agriculture
Diagnose a rice leaf from a photo
Show the lesion and get what it resembles, what it is commonly confused with, and which provincial extension office to call — with an explicit refusal to prescribe a chemical it cannot be sure about.
Read a fertiliser bag and do the arithmetic
Photograph the bag, say the plot size out loud in rai, and get the mixing rate back as a spoken answer, offline, standing in the field.
Commerce and finance
Turn a pile of receipts into a VAT table
Photograph the stack. Get a table with vendor, tax ID, date, VAT and total, flagged where a receipt is not a full tax invoice and will not survive an audit.
Reconcile a PromptPay slip against an invoice
A screenshot of a transfer slip, matched to the invoice it pays, with the amount and reference checked and the mismatch named rather than silently ignored.
Count a shelf
Photograph the shelf and get a Thai product list with facings and gaps — the job that today is a person with a clipboard walking an aisle twice.
Education
Build a lesson from a textbook page
Photograph the page and get a lesson plan mapped to the national curriculum outcomes, with the two outcomes the page does not cover named explicitly.
Mark handwritten Thai homework
Photograph a student's page and get spelling, tone-mark and spacing errors flagged with the rule each one breaks — so the marking is teaching rather than scoring.
Tourism and services
Translate a menu without losing the dish
Photograph the menu. The guest gets their language and the allergens; the dish keeps its Thai name, because ผัดกะเพรา is not “stir-fried basil” and a restaurant should not have to pretend it is.
Read an inscription
Photograph a temple inscription and get a transliteration, a plain-Thai reading, and an honest note about which characters are worn past legibility.
Industry and logistics
Put a delivery note into the system
Photograph the docket at the loading bay and have the structured fields land in the WMS — including the handwritten quantity correction somebody made in pen.
Write the incident report from the frame
A CCTV still becomes a described, timestamped incident note in Thai, in the register the report has to be filed in, for a human to sign.
Accessibility
Describe the room
Point the camera and hear what is in front of you, in Thai, at walking speed — the version of this technology that is not a productivity feature at all.
What it looks like in the API
The same endpoint. An image in the content array, and the router does the rest — which is the whole reason routing shipped first.
POST /v1/messages
{
"model": "auto", // the router picks
"messages": [{
"role": "user",
"content": [
{ "type": "image", "source": { "type": "base64", "media_type": "image/jpeg", "data": "…" } },
{ "type": "text", "text": "แบบฟอร์มนี้ยังขาดช่องไหนบ้าง" }
]
}],
"thai": { "register": "official", "dialect": "auto" }
}
// 200 OK
// { "routed_to": "ThaiLLM-Vision", "then": "ThaiLLM-Base",
// "region": "th-bkk-1", "usage": { "baht": 0.42 } }Roadmap
Four phases. None of these are dates: each one opens when the evaluation suite for it passes, and the suite is graded by Thai speakers rather than by a leaderboard.
- 1Phase 1 · Document vision
Printed and scanned Thai documents into structured fields. Buddhist Era and Thai numerals handled at the reader, not patched afterwards.
Gate - 2Phase 2 · Photograph and handwriting
Phone photos at an angle, stamps, seals, and Thai handwriting. This is the phase that decides whether the product is useful outside an office.
Gate - 3Phase 3 · Voice, all four varieties
Speech in and speech out, graded separately for Central Thai, Isan, Kam Mueang and Pak Tai — because an average across them hides exactly the failure that matters.
Gate - 4Phase 4 · Cross-modal on device
A compressed vision model that runs on a mid-range Android phone with no connection, for the work that happens where there is no signal.
Gate
References
Background reading on multimodal models, from primary documentation and the papers the field is built on. External links, not affiliated with ThaiGPT+.
IBMWhat is a multimodal LLM? — IBM
The clearest short definition of a multimodal large language model and what it is for.
NVIDIAWhat are vision-language models? — NVIDIA glossaryHow an image encoder is joined to a language model to give it sight.
WikipediaVision-language model — WikipediaA neutral overview with a wide bibliography, if you want the shape of the field before the detail.
AnthropicVision — Claude API documentationWorking documentation for passing images into a model, and the practical limits of doing so.
GoogleImage understanding — Gemini API documentationA natively multimodal API, including document and video input.
OpenAIImages and vision — OpenAI platform documentationThe other widely used vision API, for comparison of request shape.
Hugging FaceImage-text-to-text — Hugging FaceThe open-model side: task definition, and models you can run yourself.
arXiv 2103.00020Learning Transferable Visual Models From Natural Language SupervisionCLIP, 2021 — image and text trained into a shared representation. The paper the modern field starts from.
arXiv 2204.14198Flamingo: a Visual Language Model for Few-Shot Learning2022 — how a frozen language model is taught to attend to images.
arXiv 2304.08485Visual Instruction Tuning2023 — LLaVA, and the step that made vision-language models conversational.
ໂຄງການກ່ອນເປີດຕົວ
ເລືອກວິທີເຂົ້າຮ່ວມ
ເປີດຮັບສາມກຸ່ມ. ເລືອກກຸ່ມໜຶ່ງແລ້ວ ແບບຟອມຂ້າງລຸ່ມຈະຖືກຕັ້ງໃຫ້ — ປ່ຽນໄດ້ຢູ່ບ່ອນນັ້ນ.
ເປີດລົງທະບຽນລ່ວງໜ້າ
ລົງທະບຽນລ່ວງໜ້າ
ສາມຊ່ອງ. ບອກພວກເຮົາວ່າທ່ານແມ່ນໃຜ ແລະ ຈະຕິດຕໍ່ໄປໃສ ແລ້ວພວກເຮົາຈະສົ່ງອີເມວເມື່ອຄຳເຊີນຂອງທ່ານພ້ອມ.
- ສາມຊ່ອງ, ບໍ່ຕ້ອງຄົ້ນຫາຫຍັງ
- ສອງການຍິນຍອມ, ແຍກຈາກກັນ, ບໍ່ໄດ້ຕິກໄວ້ລ່ວງໜ້າ
- ໄດ້ຮັບລຳດັບຄິວທັນທີ
ໃຊ້ໄດ້ເທິງ 3G · ภาษาไทย / English