预发布 · 报名开放
ปัญญาประดิษฐ์ ภาษาไทย AI that works in Thai, priced in baht.
ThaiGPT+ runs on Thai models, hosted in Thailand, and improved by Thai contributors. Drafting, document Q&A with citations, translation, dialect and voice — on a phone, on a desktop, or through an API you can call from anything. The + is what comes next: vision, hearing, and a router that picks the right model for you.
选一个问题,看看 ThaiGPT+ 如何作答。
产品
Four surfaces, one brand.
A mobile app built for a phone-first market with the highest mobile penetration in Southeast Asia. A desktop workspace built for institutions. An API built for everyone else. And the training loop that keeps improving what sits behind all three. These are live, interactive previews — click through them.
移动端——语音优先,断网也能用
Thailand has around 135 mobile connections for every 100 people. The phone is not a secondary surface here, it is the surface. The microphone is the primary control: Thai keyboard input is slow, and speech reaches users who read slowly or read a different dialect than they speak.
网络随便切。你看到的是应用,不是错误页。
Thailand is phone-first by a wider margin than almost anywhere: more mobile connections than people, and most of the country reaching the internet through one. The app is designed for the connection people actually have, and it keeps working when they have none.
- 完整模型,完整上下文
- 带出处的文档问答
- 语音输入与语音输出
- Live translation TH ↔ EN ↔ ZH
每次回答都按泰铢计量,并显示由哪个区域处理。
用泰语作答。
ผมไม่มีราคาตลาดแบบเรียลไทม์
I don't have live market prices. I can show this week's ministry bulletin, or draft a message to your co-op.
草稿已生成——46 个词,通俗语体。点按可在发送前修改。
点按麦克风——语音是主要操作方式
The models
One family, six jobs.
A single Thai model is the wrong shape for a country where the same platform has to answer a ministry's 400-page regulatory corpus and a farmer's question on a phone with no signal. ThaiGPT+ ships a family, and the router picks between them. Sizes and context windows are targets for the pre-release, not benchmark results.
Four ways to say the same thing
This is the part that has no Lao equivalent and no off-the-shelf answer. Thai is not one language in practice: around a third of the country speaks a regional variety at home and writes Central Thai at work, and a model that only knows the written standard cannot hear half its users.
The written standard. What every general-purpose model has seen, and roughly all it has seen.
The largest regional variety, spoken across the northeast and close enough to Lao that the work crosses the border in both directions.
Northern Thai, with its own historic script still in liturgical and heritage use.
Southern Thai — fast, heavily contracted, and the variety speech systems get wrong first.
And five ways to be polite about it
Register in Thai is not a tone setting, it is a different vocabulary. Getting it wrong in a letter to a ministry is not a style error, it is a social one — and it is the single thing a foreign model is most reliably bad at.
- 1ราชาศัพท์ · royal
- 2ภาษาราชการ · official
- 3ทางการ · formal
- 4สุภาพ · polite
- 5ภาษาพูด · plain
Register is a request parameter, not a prompt instruction — so it is auditable, and a workspace can pin it.
See how the router picksSmart routing
One endpoint. The right model, every request.
Most requests do not need the biggest model, and some must never leave the phone. ThaiGPT+ reads each request before a token is generated, and decides which model answers it, which pipeline it goes through, and which region it is allowed to run in. You call one endpoint and never pick a model again.
How the router worksComing soon Roadmap — not yet shipped
Next: it learns to see, read and hear Thai.
Text was the first sense. ThaiGPT+ Vision, Scribe and Voice add the others — a model that can be shown a photographed form, a stamped and scanned circular, a handwritten minute, a shelf, a leaf or a prescription label, and answer about it in Thai.
ThaiLLM-Vision
Photographs, screenshots, charts, shelves, scenes, forms. Answers about what is in the frame, in Thai, with the region it was served from.
ThaiLLM-Scribe
The document specialist: stamped, scanned, skewed, multi-column, part-handwritten Thai paperwork into structured fields you can put in a database.
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.
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.
平台
PrivateGPT-Asia 的全部能力。
ThaiGPT+ 继承了一套完整的私有 AI 应用层,而不是重新搭建。下面列出的都是平台自带的能力,第一天即可使用,并在境内运行。
带出处的检索
可以追溯到具体页码与段落的回答。打不开来核对的说法,就是用不上的说法。
文档接入
PDFs, scans, spreadsheets and office documents — including Thai documents that were typed, printed, stamped, then scanned again.
工具与 MCP
模型通过 Model Context Protocol 调用贵单位的系统,而不是猜测里面有什么。
技能
为组织中反复出现的工作预先做好的流程:一份通知、一种报告格式、一张核对清单。
Text-to-SQL
基于贵单位自己的数据表回答问题,并把查询语句显示出来,便于有人核验。
即时的 API 兼容
与 Claude API 同构的接口。现有客户端只需改一个基础 URL,就指向泰国境内的端点。
放在哪里就在哪里运行
境内服务器、贵单位自有场地,或者完全断网。平台不要求绕经境外服务商。
没有东西离开所在场地
私密来自架构而非条款:数据留在部署平台的那个环境里。
以及只有这里才有的部分
The platform gives ThaiGPT+ everything above on day one. Eight endpoints exist because Thai needs them and no general-purpose vendor is going to build them: word segmentation for text written without spaces, normalisation of Buddhist Era dates and Thai numerals, translation with register control, transliteration to and from RTGS, speech, dialect conversion, honorific grading, and Thai document OCR.
为何成本更低
在外国模型里,说泰语更贵。
Thai is written without spaces between words, is tonal, stacks up to four levels of vowel and tone mark around a single consonant, and carries heavy Pali-Sanskrit vocabulary. General-purpose tokenizers fragment it badly — often to the character — so a Thai user pays a multiple of the English price per sentence and gets a worse answer back. A Thai-optimised tokenizer attacks both sides of that ratio.
สรุปเอกสารนี้เป็นภาษาไทย
“summarise this document in Thai”
Thai is written without spaces between words, so a tokenizer trained on spaced languages falls back to characters — and a stacked vowel or tone mark often becomes a token of its own. You pay per fragment.
Trained on Thai, it segments close to real word boundaries and keeps a stacked cluster whole — fewer tokens in, more context per request, and a better answer out.
放到实际用量上是这样
拖动滑块,设定贵单位办公室或应用每天真正处理的泰语量。词元数按上面那句话换算。
Late-2025 measurements, DataReportal Digital 2026: Thailand. Used for product sizing.
数据驻留与同意
可以验证,而不只是声称。
“主权”两个字谁都能写在幻灯片上。以下四处是 ThaiGPT+ 让它可以被外部核验的地方——写在产品里,不是写在政策文件里。
The residency badge is one of the three places the kranok signature is used at all. It stays in permanent view in the institutional workspace.
"region": "th-bkk-1"每个 API 响应都带着处理它的区域。数据存放地不再是宣传册上的说法,而是贵单位自有系统可以核验的一个字段。
off unless you turn it on产品动态与评估数据使用是两个彼此独立的决定。两者都不会预先勾选,都可以随时更改,而撤回会一路贯穿到处理链路的末端。
logged · exportableWorkspace activity is logged and exportable, so an institution can answer its own oversight questions — including under the PDPA — without asking us for a report.
egress: blockedAn institutional workspace can pin itself to a region, and the router declines rather than reaches for a model outside it. On-premise and air-gapped deployments are supported by the platform.
三方合作
三家公司,一个平台。
ThaiGPT+ 既不是换层外壳,也不是分销协议。每一方各自负责一层,产品只存在于三者交汇之处。
XTND Thailand brings ThaiGPT+ to market. It owns the relationship with the people and institutions who use it — ministries, universities, hospitals, banks, schools and the public — and it carries the commercial model: baht-denominated pricing, prepaid credit through PromptPay and local bank rails, and free access for education.
- Market entry, distribution and institutional onboarding
- Baht pricing, prepaid credit, PromptPay and local bank rails
- The Thai contributor programme behind the training loop
XTND Dynamics builds the product. Part of XTND Group, it takes the PrivateGPT-Asia platform and turns it into four shipping surfaces — a mobile app, a Windows and web workspace, a developer API and a training platform — then engineers the Thai layer that no general-purpose platform provides, including the router and the multimodal pipeline.
- 平台之上的移动端、桌面端与网页应用
- Thai segmentation, normalisation, register, dialect and speech services
- The router, the multimodal pipeline and the consent ledger
PrivateGPT-Asia 是底层平台,也是 PrivateGPT 在本地区的落点。它是私有 AI 的完整应用层——带出处的检索、文档接入、工具、MCP、技能与 text-to-SQL——统一置于与 Claude API 同构的接口之后,可在境内、贵单位自有场地或完全断网的环境中运行,数据不出所在环境。
- 检索、接入以及可追溯到出处的回答
- 面向贵单位自有数据的工具、MCP、技能与 text-to-SQL
- 即时的 API 兼容与断网环境部署
各方分别负责什么
把鼠标移到上方的合作方或下方的层级上,即可看到各自的工作落在哪里。ThaiGPT+ 是商业平台层——它有意既不拥有模型,也不拥有数据中心。
应用平台是 PrivateGPT-Asia——与本地区在用的同一套私有 AI 层,为泰国在境内运行。
预发布计划
选择你的加入方式。
三类名额已开放。选择其一即会配置下方表单——你也可以在那里更改。
预发布 · 报名开放
加入预发布。
三个字段。告诉我们你是谁、往哪里联系,邀请准备好后我们会发邮件通知。
- 三个字段,无需查找任何资料
- 两项同意,彼此独立,均未预先勾选
- 立即返回排队位置
3G 网络下也能用 · ภาษาไทย / English