ThaiLLM: what it is, what it runs on, where it is hosted

Research note, October 27, 2026, by Alexander Grabowski (B4L). Sources were retrieved on October 27, 2026. Figures are cited, and our own estimates are marked.

This note covers ThaiLLM: who runs it, which models it has released, what they were trained on, what that training probably cost and where the models are served. Money is in THB, with USD in parentheses at 33 THB per USD, the B4L planning rate.

In short

• ThaiLLM is Thailand's national program for open-weight Thai language models. The Big Data Institute (BDI, under the Ministry of Digital Economy and Society, MDES) leads it with NECTEC, VISTEC, AIEAT and AIAT, and the Digital Economy and Society Development Fund (DE Fund) pays for it.[30][31]
• It has released two foundation models, ThaiLLM-8B and ThaiLLM-30B. Both are Alibaba Qwen3 models with continued pre-training on about 63 billion tokens, and both are licensed under Apache-2.0.[1][2][32]
• Training ran on LANTA, NSTDA's supercomputer with 704 NVIDIA A100 GPUs, in service since November 2022.[3][4] The GPU-hours used for ThaiLLM are not published. We estimate the final training runs of both models at about 9,000 to 23,000 A100-hours, which comes to roughly 0.1 to 1.5 million THB depending on the price, and to 0.2 to 4 million THB with experiments at LANTA's rates.
• The free Playground and the OpenAI-compatible API run on compute from National Telecom (NT) and SIAM.AI CLOUD rather than on LANTA.[5][27][33] On October 27, 2026, the API listed six models: four 8B ThaiLLM derivatives and two stock Qwen models.[34]
• Phase 1 cost about 80 million THB (USD 2.4 million).[31] Phase 2 has a framework of about 100 million THB (USD 3.0 million), approved year by year.[31]
• We found no role for Chiang Mai University (CMU) or any other Chiang Mai organization. The published partner list names Mahidol and Chulalongkorn universities but not CMU.[3]
• In our reading, every model ThaiLLM serves fits on one or two data-center GPUs (our estimate). ThaiLLM therefore shows what to host in Chiang Mai, and nothing about its hardware stands in the way.

Who runs it

Lead and partners. BDI, a public organization under MDES, is the lead agency[31] and the legal service provider of the Playground and API.[27] The official site names BDI, NECTEC, VISTEC, AIEAT and AIAT as the public-private partners.[30] NSTDA's list is longer: NECTEC/NSTDA (under the Ministry of Higher Education, Science, Research and Innovation, MHESI) with BDI, VISTEC, AIEAT, AIAT, Mahidol University and Chulalongkorn University.[3] Part of the training data came from a "National Data Bank," to which the National Research Council of Thailand, the Office of the Council of State and the National Library contributed.[35]

Funding. The money comes from the DE Fund of MDES.[30][36] The budget figures differ by source and date:

We take 80 million THB as the figure for Phase 1, because the BDI director gave it after the launch and the Bangkok Post agrees. The 89 million THB was a budget set before launch, and the 120 million THB a plan from 2024. We do not know whether the 100 million THB framework for Phase 2 was approved, or how it relates to the 52 million THB line.

Timeline. The ThaiLLM-30B repository on Hugging Face was created on August 9, 2025, and the 8B on December 3, 2025.[32] The partners opened both models to the public in January 2026.[3] MDES and MHESI held the official launch in Bangkok on May 5, 2026, when BDI reported more than 8,000 Playground users.[5] Post Today later reported more than 10,000.[40]

Chiang Mai. No partner list we found includes CMU.[3][30] MHESI says researchers from 150 universities and six research institutes help refine the datasets, but it names none of them.[35]

The models

ThaiLLM has no architecture of its own. ThaiLLM-8B uses Qwen's Qwen3ForCausalLM code path, and ThaiLLM-30B is a Qwen3 mixture-of-experts model (MoE, in which each token uses only a few "expert" sub-networks).[41][42] Both are continued pre-training (CPT, further training of an existing model on new text) of Qwen3 base models, and both model cards say they are base models that need instruction tuning before use.[1][2]

The table draws on the model configs,[41][42][43] the Hugging Face listing of the ThaiLLM organization (dates, tags and bases)[32] and the cards of the 30B-IQ and Prescreen models.[44][45] For the variants, "Released" is the date the repository was created. The Prescreen model is not based on Qwen at all; it is a fine-tune of Google's MedGemma.[45]

Corpus. The 8B card itemizes the training corpus: 24.0B English tokens from FineWeb2, 31.5B Thai tokens from FineWeb2 and 8.05B curated tokens.[1] By our arithmetic that sums to 63.6B tokens, and the 30B card gives the same corpus of about 63 billion tokens.[2] Other sources disagree. NSTDA's English page says ThaiLLM "has been trained on over 100 billion tokens of high-quality Thai data,"[36] while two partner papers give 63B and 64B Thai tokens.[46][47] We trust the model card, because it lists the datasets and their token counts. On the card, at most about 40B of the tokens are Thai (our estimate: 31.5B plus 8.05B, if all the curated data is Thai), and 24B are English.

Benchmarks (self-reported). On its own card, ThaiLLM-8B raises the ThaiExam average from 0.314 for Qwen3-8B-Base to 0.483 and the suite average from 0.5987 to 0.6891, while English MMLU drops slightly (0.7691 to 0.7565).[1] ThaiLLM-30B raises ThaiExam from 0.5947 to 0.6478.[2]

The partner models built on ThaiLLM

The base models are not meant for chat. Four partner groups fine-tune them, and their instruct models are what the Playground serves.[30][34] The note on Thai models gives more detail and comparisons.

• OpenThaiGPT-ThaiLLM-8B-Instruct v0.7.2 comes from AIEAT with iApp Technology (the project now calls itself OpenThai). It is Apache-2.0, and the API lists it as "v7.2."[48][49][34]
• Typhoon-S-ThaiLLM-8B-Instruct (SCB 10X) has a 32K context and an Apache 2.0 license.[50] On SCB 10X's Thai-only suite it averages 71.20 against 66.66 for Qwen3-8B (self-reported).[47]
• Pathumma-ThaiLLM-8B think 3.0.0 (NECTEC) is a supervised fine-tune of the ThaiLLM base; the evidence does not state its license.[51] A newer 4B Pathumma 4.0.0, built on a 4B ThaiLLM base that we found no separate release of (262,144-token context, Apache-2.0), is not on the API.[52]
• THaLLE-0.2-ThaiLLM-8B-fa (KBTG) is a mergekit merge of ThaiLLM-8B with Qwen3-8B and THaLLE, KBTG's finance model, under Apache-2.0.[53][54]

Training hardware

LANTA. BDI and NSTDA both say the ThaiLLM models were trained on LANTA, the supercomputer of ThaiSC (NECTEC/NSTDA).[5][3][55] LANTA is a 346-node HPE Cray EX cluster with 176 GPU nodes holding 704 NVIDIA A100 GPUs, in service since November 2022.[56] Each GPU node holds four A100s and one 64-core AMD EPYC CPU.[56] TOP500 lists the GPUs as A100 40GB, with a Linpack result of 8.15 PFlop/s against a theoretical peak of 13.77 and a power draw of 310.46 kW.[57] Its rank fell from 70th in November 2022 to 223rd in June 2026 with the same figures on every list, so no hardware appears to have been added.[57]

ThaiSC's address is the Innovation Cluster 2 building at Thailand Science Park, Khlong Luang, Pathum Thani.[56] The page does not say where the machine room is. NSTDA claims that LANTA's liquid cooling uses 33% less cooling energy than air cooling, a first for Thailand.[55]

Public and private organizations can rent LANTA.[55] A Thai company pays 45 THB per Service Node-Hour, and a GPU node-hour (four A100s) counts as 3 of them, so it costs 135 THB (USD 4.1).[58] By our arithmetic, that is about 34 THB (USD 1.0) per A100-hour. Projects start at 5,000 THB (USD 150).[55]

GPU-hours. No technical report exists, and the GPU count, GPU-hours, duration and software of the ThaiLLM-8B and 30B runs are unpublished. Two related figures are published:

• ThaiSC delivered 3.18 million GPU-hours to 730 users from 69 organizations between October 1, 2025, and June 30, 2026.[59] That is about 69% of LANTA's GPU capacity (our estimate: 704 GPUs x 6,552 hours = 4.61 million GPU-hours). ThaiLLM's share is not given.
• NSTDA offers an illustration: training an 8B model from scratch would take about 87 days on about 700 LANTA GPUs.[55] That is roughly 1.46 million GPU-hours (our arithmetic: 87 x 24 x 700). It does not describe the ThaiLLM run, which continued from Qwen3 on a much smaller corpus.

What the training probably cost. Since no figures are published, we estimate the compute from the model size and the corpus. Training takes about 6 floating-point operations per parameter per token,[306] and an A100 delivers at most 312 TFLOPS in BF16.[305] We assume 35 to 50% of that peak for the dense 8B model and 15 to 30% for the 30B mixture-of-experts model (about 3 billion parameters active per token), and we add up to 50% for long sequences and restarts. We then price the GPU-hours at LANTA's published rates (11.25 THB per A100-hour for government and education, 33.75 THB for a Thai company)[58] and at Lambda's on-demand price for an A100 40GB of USD 1.99 per GPU-hour.[23]

All figures in the table are our estimates. On all 704 LANTA GPUs the two final runs would take between half a day and a day and a half; on a 64-GPU allocation like Pathumma's, one to two weeks. Experiments, failed runs and evaluation typically multiply the final run by two to five (our judgment), which puts ThaiLLM's training compute at roughly 0.2 to 4 million THB at LANTA's rates, or up to about 8 million THB at cloud prices. If NSTDA's "over 100 billion tokens" is right, the figures rise by about 60%. Either way, the GPUs were a small part of the 80 million THB spent on Phase 1. In our reading, most of the money went to data, people, evaluation and the hosted platform, though no breakdown is published.

Partner fine-tunes. The partners disclose more than the core project does:

SCB 10X calls its budget "academic," equivalent to two days on a single H100 node.[50] At Siam AI's on-demand list price, from USD 2.69 an hour for an H100 (vendor price page, read October 27, 2026),[61] 8 GPUs for 48 hours would cost about USD 1,030 (34,000 THB) (our estimate: 384 GPU-hours x USD 2.69).

Where it is hosted and how it is served

Channels and access. ThaiLLM is free through three channels: a chat Playground, an API and model downloads.[3][27] The API follows the OpenAI SDK format.[33] A third-party tutorial uses the base URL https://thaillm.or.th/api/v1 with a key for each user.[62] BDI issues API tokens to individual users and forbids passing them on.[27] The only rate limit we found (5 requests per second, 200 per minute) comes from a vendor blog and is unverified.[63]

Compute. BDI says the Playground (API and chat sandbox) gets its compute for development and serving from NT and SIAM.AI CLOUD.[5] The project site credits NT for the compute on which the system was installed.[30] Before the launch, BDI said the first version was installed on NT's cloud.[38] OpenThai, a partner, also names NT and SIAM.AI Cloud.[64] A vendor blog and a directory site claim that ThaiLLM is served from LANTA.[63][65] We give that claim little weight, because BDI, the operator, names NT and SIAM.AI CLOUD.

Location. No source names the data center or the city. The site and API sit behind Cloudflare, which hides the origin server.[34] Our probe on October 27, 2026, reached Cloudflare's edges in Singapore and Bangkok, which says nothing about the origin. Siam AI's GB200 equipment is at STT Bangkok 1, and the Board of Investment approved a Siam AI cloud project in Chonburi and Pathum Thani;[66] we do not know whether either serves ThaiLLM. The privacy notice lets BDI transfer personal data to cloud servers abroad, for example in Singapore or the United States.[67] The clause is generic and may cover sign-in or analytics services rather than inference.

Models on the live API. An unauthenticated request to the models endpoint on October 27, 2026, returned these models:[34]

The national platform therefore also serves stock Alibaba models, and it lists no ThaiLLM-30B model. The inference software is not published.

Terms versus license. The Playground terms, dated January 19, 2026, differ sharply from the license of the weights:

• Outputs may not be used, modified or published for commercial purposes without BDI's prior written consent.[27]
• Users grant BDI the right to use their inputs to develop and train BDI's models, and inputs must not contain personal data.[27]
• The downloadable weights are Apache-2.0, which allows commercial use, and BDI itself names downloading and self-hosting as the route to full control of data.[32][31]

Siam AI. OBON Corporation holds more than 99.86% of Siam AI's shares (January 2025).[68] On May 8, 2026, Bloomberg, as reported by the Bangkok Post, identified OBON as "Company-1" in a US case on GPU servers allegedly diverted to China; a source said OBON is one of Siam AI's server suppliers.[69] Siam AI denied exporting AI servers to China and said it complies with US export controls.[70] These remain allegations. The note on Thai compute has the detail.

Roadmap and national plans

• Phase 2 was to start in late May 2026, with its budget approved year by year. It adds multimodal models (images, video, audio), Thai regional dialects, domain benchmarks (for example, how good a medical bot must be to meet the criteria of the Thai FDA), training and hackathons.[31] NSTDA says the partners plan larger models and a move toward agentic AI.[3]
• ThaiLLM 2027, presented at BDI Day on October 8, 2026, targets foundation models for regional dialects, extraction benchmarks and filters that mask private data.[71]
• Nemotron. NVIDIA said in September 2026 that BDI and iApp, as ThaiLLM members, are "exploring Nemotron as a foundation model."[72] iApp's OpenThai 2.0 Legal already builds on Nemotron 3 Nano (30B total, about 3B active) and carries the NVIDIA Open Model Agreement instead of Apache-2.0.[73][74]
• Revenue. BDI expects ThaiLLM to earn revenue by 2027 from custom training for agencies and companies.[38]
• National AI Strategy (NAIS). ThaiLLM is center of excellence #7, the "ThaiLLM Consortium"; #8 is a separate "Government AI Computing Center."[75] The budget framework sets at least 2,000 million THB (USD 61 million) for open-source AI infrastructure and the National Data Bank. It is approved in principle only and "may change post-government transition."[75]
• No dedicated ThaiLLM GPU cluster has been announced. The budget of the government cloud GDCC, which pays NT to run it, rises to 4,980 million THB (USD 151 million) in FY2027.[76][77] The DE minister said in August 2026 that it will become a "National Cloud" with stronger GPU infrastructure, but gave no GPU count.[78]

What this means for the investor

This section is our judgment.

ThaiLLM has no hardware advantage that a newcomer would need to match. It is a continued pre-training of Qwen3 plus partner fine-tunes, and every model it serves has 8B to 35B parameters. ThaiLLM-8B in BF16 (16-bit weights) needs about 16 GB of GPU memory and ThaiLLM-30B about 61 GB (our estimate: parameters x 2 bytes, before the memory for long contexts). A single 80 GB GPU can serve either; the note on hardware covers sizing.

Training is cheap, and serving is the steady cost. A partner's post-training run took one 8-GPU node for two days, about USD 1,000 rented (our estimate above). We recommend that the investor rent GPUs for fine-tuning, on LANTA (about 34 THB per A100-hour for a Thai company) or a GPU cloud, and own hardware only for steady inference. The notes on option A and option B make that comparison.

The public API should not become a production dependency. Outputs cannot be used commercially without consent, inputs may train BDI's models, and no rate limit or SLA is published. The Apache-2.0 weights can be self-hosted commercially, and that is the route for customer data.

ThaiLLM runs on partner clouds at undisclosed sites, and we know of nothing in Chiang Mai. NT, a ThaiLLM serving partner, lists a Tier 3 data center in Chiang Mai with 25 racks, under renovation with completion estimated for June 2026.[12] NT is also a partner of CMU's Academy of AI.[79] A Chiang Mai inference node, possibly with CMU and with Northern Thai dialect data for Phase 2, is a partnership worth proposing, although no source says BDI wants one.

The investor should plan for model churn. The base may move from Qwen to Nemotron, which carries a different license, so the hardware should not be tuned to one model family.

One partner carries legal risk. Press reports about an export-control case name Siam AI's majority owner. The investor should keep its hardware supply chain documented, from vendor to serial number.

Open questions

• How many GPUs, GPU-hours and days did the training of ThaiLLM-8B and 30B take, and with which software? Ask ThaiSC (NECTEC/NSTDA) or BDI.
• Which data centers, GPUs and inference software serve the Playground and API, how is the load split between NT and SIAM.AI CLOUD, and what are the rate limits? Ask the BDI ThaiLLM team.
• Would BDI take an inference-hosting partner in Chiang Mai, as a regional node or for work on Northern Thai dialects? Ask BDI, and ask NT about its Chiang Mai data center after June 2026.
• Why does NSTDA say over 100 billion tokens when the model card says about 63 billion? Ask BDI or NECTEC.
• Was the 100 million THB framework for Phase 2 approved, and how does it relate to the 52 million THB line for FY2027? Ask BDI.
• Will the Phase 2 models stay Apache-2.0, and will the ban on commercial use of API outputs change with BDI's revenue plan for 2027? Ask BDI.
• Is CMU among the 150 universities refining ThaiLLM's datasets? Ask MHESI or CMU.