Thai-capable open models: what the investor could host

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 the open models that handle Thai and that the investor could host: who makes them, how they were trained, how good they are at Thai, what their licenses allow and what hardware serves them.

In short

• Every Thai model released in 2025 and 2026 continues or fine-tunes a foreign open model (Qwen, Llama, Gemma or NVIDIA Nemotron). Even the national ThaiLLM-8B is Qwen3-8B-Base trained further on about 63 billion tokens.[1][47]
• Most Thai model work fits on one 8-GPU server. SCB 10X trained all Typhoon 2 sizes, up to 70B, on a single node of eight H100 GPUs, and Jasmine trained OpenJAI on 8x H100 in under a day.[6][7]
• Strong general models already handle Thai well. In OpenThai's own table, its plain Qwen3.8-27B base scores 0.820 on OpenThaiEval (Thai national exams), ahead of Typhoon 2.5 at 0.742, and the Thai fine-tune adds 0.022 to reach 0.842.[80]
• Most weights may be hosted commercially, under Apache 2.0, MIT, Gemma, Llama or NVIDIA licenses. Typhoon is the exception to check: its terms say they apply "regardless of the method of access," and they forbid training on outputs and competitive benchmarking.[81]
• As of October 2026, no maker sells a paid, production API hosted in Thailand. Typhoon's free beta allows 5 requests per second and 200 per minute, and its AWS listing still reads "Coming soon"; SEA-LION's free API allows 10 requests per minute.[9][82]
• Serving needs are modest. Typhoon 2.5 (30B), OpenThai 2.0 (27B) and gpt-oss-120b each run on one 80 GB GPU; only frontier models such as DeepSeek V3 (671B) need a full 8-GPU H200 or B200 server.[8][80][83]

The landscape in one table

Sizes are given as total parameters / parameters active per token. Mixture-of-experts (MoE) models use only a fraction of their weights for each token, which makes them fast for their size. Memory figures marked est. are our estimates of the weights alone: parameters x 2 bytes at BF16, x 1 byte at 8-bit and about x 0.5 at 4-bit, before the working memory that each conversation needs (the KV cache). Prices are converted at the B4L planning rate of 33 THB per USD.

How they were trained

Nobody in Thailand trains a language model from scratch. The national base models are continued pretraining of Alibaba's Qwen3 on about 63 billion tokens: 31.5 billion of Thai web text, 24 billion of English web text and 8 billion of curated data such as research articles and news.[1][2] The company models fine-tune either ThaiLLM or a foreign base directly. Typhoon 2.5 and OpenJAI start from Qwen3, and WangchanLION-v3 from Llama 3.1.[88][7][95] SEA-LION v4.8 and OpenThai 2.0 Legal build on NVIDIA Nemotron.[99][73] NSTDA's own illustration shows why: training an 8B model from scratch would take two decades on an eight-GPU workstation, or about 87 days on LANTA's roughly 700 GPUs.[55]

Typhoon 2.5's own training compute is not published. Together AI says in a vendor case study that SCB 10X used its GPU clusters and serverless inference to train and deploy Typhoon models, so at least part of Typhoon's work ran on rented cloud.[122]

The pattern is clear. Apart from the work on the LANTA supercomputer (ThaiLLM's continued pretraining and NECTEC's Pathumma post-training) and AI Singapore's regional pretraining on 72 H100s, every team above did its work on one 8-GPU node or less. Even the continued pretraining of a 70B model (Typhoon 2) fit on one node. We read NECTEC's use of 64 A100s as taking what LANTA offers rather than as a need, because SCB 10X post-trained the same ThaiLLM-8B base on 8 H100s in about two days.[47] Most teams rented or borrowed their GPUs (Siam AI donated OpenThai's), and Jasmine is the only team in our sources that trained on its own in-house GPUs. We conclude that a single 8-GPU server, of the kind the investor is considering, matches the scale at which Thai model work is actually done.

How good they are at Thai

All the published comparisons below are self-reported by the makers and run on their own test setups, and they often compare models of different sizes. We found no neutral leaderboard for 2026 that covers all these families on one test setup. The numbers show a direction; they do not rank the models.

The general base already scores high. On OpenThaiEval (Thai national exams), OpenThai 2.0 scores 0.842, its plain Qwen3.8-27B base 0.820, Typhoon 2.5 0.742 and Pathumma think 3.0 (8B) 0.660, all measured by iApp.[80] The Thai fine-tune adds two points over the general model it came from. Typhoon 2.5 uses only about 3B active parameters, so the comparison is not like for like.

A general model beat the Thai fine-tunes on Thai knowledge. In Jasmine's table, plain Qwen3-14B scores 57.5 on the Thai MMLU-ProX test, ahead of OpenJAI (54.7), OpenThaiGPT 1.5 14B (49.3), Typhoon 2.1 12B (45.2) and GPT-4.1-nano (39.8).[7] OpenJAI did edge out Qwen3-14B on Thai instruction following (39.4 vs 38.1), and on Thai chat (MT-Bench) all of them score about 8.0 to 8.1 except OpenThaiGPT 1.5 at 6.9.[7]

Fine-tunes win on Thai but sometimes lose overall. Typhoon-S-8B beats Qwen3-8B on SCB 10X's Thai-only suite, 71.20 vs 66.66, and on Thai chat, 7.89 vs 7.08. Across the full English and Thai suite, however, Qwen3-8B averages 54.02 against Typhoon-S's 49.99, and Qwen3-8B also follows Thai instructions better (80.47 vs 76.45).[47]

Domain fine-tunes help in their domain. KBTG's THaLLE-0.2 scores 0.779 on the Thai M3 Exam in thinking mode, against 0.706 for Qwen3-8B and 0.720 for ThaiLLM-8B-Instruct.[53] OpenThai 2.0 Legal recalls sections of the Thai Revenue Code better than Qwen3.6-35B (0.40 vs 0.36), but when the law is supplied in the prompt, both score 0.99 at citing the applicable sections of the Civil and Commercial Code.[73]

The national base adds Thai skill, mostly at the small size. ThaiLLM-8B lifts the ThaiExam average from 0.314 to 0.483 over Qwen3-8B-Base, while English MMLU dips from 0.769 to 0.757.[1] ThaiLLM-30B gains less over its Qwen base: 0.595 to 0.648.[2]

Specialists matter for reading and speech. On Thai handwriting, SCB 10X's 2B Typhoon OCR 1.5 has a character error rate of 0.054, against 0.261 for the 27B OpenThai 2.0, in iApp's own test.[80]

Vendors make larger headline claims. SCB 10X says Typhoon 2.5 "Matches GPT-4o and Claude Sonnet 4 in benchmarks," but it publishes the tables only as images and says the model is "Not yet for deep reasoning."[88] AI Singapore says Qwen-SEA-LION v4 ranked 6th of 59 on its SEA-HELM benchmark, which includes a Thai exam, and first among open models under 200B at release.[123]

In our reading, a strong recent general open model (Qwen3.x class) is already good at Thai, and on some tests better than Thai fine-tunes of similar size. Thai fine-tunes add value where the task is specifically Thai: natural Thai phrasing and culture, Thai law, finance and medicine, and above all handwriting OCR and speech recognition, where small specialist models beat large general ones. The investor should check this on the workloads it intends to serve before choosing.

Licenses and terms that matter for a commercial host

Most Thai and general weights allow a company to host them and sell the service (license texts as read in October 2026):

• Apache 2.0 (no user cap, attribution only) covers ThaiLLM-8B and 30B, OpenThai 2.0, Pathumma 4.0, THaLLE-0.2, OpenJAI, the Qwen3 models and gpt-oss-120b; the table above gives each source.[32][80][114]
• MIT covers SEA-LION v4.5 Qwen and v4.8 Nemotron. AI Singapore says SEA-LION is "free to use, including for commercial purposes."[101][99][102]
• The Llama 3.1 Community License (WangchanLION-v3) requires a separate license from Meta above 700 million monthly active users, and products must credit "Built with Llama."[97]
• The NVIDIA Open Model Agreement (OpenThai 2.0 Legal) says "Works are commercially usable," and derivatives may be distributed.[124]
• The Gemma terms (Typhoon 2.1, Pathumma's Gemma-27B, Gemma-SEA-LION) allow use and redistribution, but Google's use restrictions must be written into any agreement with downstream users.[112]
• User thresholds apply in two licenses. The SeaLLMs license and the Qwen license used by OpenThaiGPT 1.6 both require a separate license above 100 million monthly active users, far above anything a Chiang Mai service would reach.[105][125]

Typhoon needs a legal review. The Typhoon 2.5 weights are tagged Apache 2.0, but the model card adds that "By using this model, you agree to the OpenTyphoon Terms and Conditions."[89] Those terms, operated by SCB DataX Company Limited, acknowledge that models may be distributed open source and say the terms "apply to and govern the use of the Services regardless of the method of access."[81] They forbid using the Services or Output "to train, fine tune, or otherwise improve any Artificial Intelligence Models," forbid "competitive benchmarking, data scraping, or model extraction" without written consent, and are governed by Thai law in Thai courts.[81] Pulling the other way, SCB 10X's FAQ says it "claims no rights in Outputs you generate using Typhoon."[126] If the terms bind self-hosted weights, a host could not use Typhoon to generate training data for its own models and could not publish comparisons against it. We recommend that the investor have a Thai lawyer review this before any product depends on Typhoon.

ThaiLLM's weights and its Playground carry different terms. The downloadable ThaiLLM weights are Apache 2.0, and BDI calls them "open license" and open-weight.[5] The free Playground and API are different: outputs may not be used commercially without BDI's prior written consent, inputs may be used to train BDI's models, and inputs must not contain personal data.[27] A commercial service must therefore self-host the weights rather than resell the API.

Hosted APIs today

Before the paid Typhoon API ended, it cost about 6.6 THB (USD 0.20) per million tokens for the 12B model and about 29 THB (USD 0.88) for the 70B model.[90] For production use, SCB 10X now points customers to Float16's hosted service or to self-hosting on rented GPUs.[90] Float16 lists H100s in its Thailand region at about 143 THB (USD 4.32) an hour.[129] Thai AI Cloud advertises a Bangkok inference API "From $0.09 per million tokens" (about 3 THB) with prompts kept in Thailand, a vendor claim that does not name the models.[130]

For the investor, the point is that no maker of a Thai model offers a paid, production-grade API hosted in Thailand as of October 2026. Third-party GPU clouds fill part of the gap, and that gap is where a Chiang Mai host would compete.

Hardware to serve them

Memory decides what fits on one GPU. The table gives the weights only; the conversation cache needs room on top.

Throughput is also within reach of one server. SCB 10X says one H100 serves Typhoon 2.5 at more than 3,000 tokens per second across 64 concurrent requests (a vendor claim), which works out to about 47 tokens per second per request (our arithmetic, 3,000 / 64).[88] NVIDIA's TensorRT-LLM tables show Qwen3-30B-A3B, the architecture under Typhoon 2.5 and ThaiLLM-30B, at 9,938 output tokens per second on one RTX PRO 6000 and 26,971 on one B200; these are offline ceilings and say little about interactive service.[132] In an independent test, gpt-oss-120b on one RTX PRO 6000 handled 50 concurrent prompts at about 32 milliseconds per token, roughly 30 tokens per second per user.[131] The note on hardware covers server options, prices and power.

What this means for the investor

This section gives our judgment and advice; B4L plans and researches on the investor's behalf.

1. Host a portfolio of models. The national ThaiLLM API already works this way, with four Thai 8B fine-tunes next to two stock Qwen models.[34] We recommend that the investor plan for one strong general model (gpt-oss-120b, or a Qwen3.x MoE of 30B to 35B) for reasoning, code and English, plus Thai specialists: OpenThai 2.0 27B for Thai documents and knowledge, OpenThai 2.0 Legal for law, Typhoon 2.5 for fluent Thai chat and tool use (after the legal review), and Typhoon OCR and ASR for handwriting and speech.
2. Size it at about one GPU per large model. gpt-oss-120b (about 61 GB), Typhoon 2.5 (about 61 GB at BF16), OpenThai 2.0 at 8-bit (a 40 GB GPU) and OpenThai 2.0 Legal at 4-bit (24 GB), sharing a GPU with Typhoon OCR 1.5 (8 to 12 GB), come to about 200 GB of weights on four GPUs of 80 to 96 GB (est.).[8] That is half of an 8-GPU server. The other half can host a larger general model or run fine-tuning, which is exactly the work every Thai team has done on one 8-GPU node.
3. Prefer clean licenses for anything sold. We recommend that the investor build paid services on Apache 2.0 or MIT models (ThaiLLM, OpenThai 2.0, Qwen, gpt-oss, SEA-LION) by default. The CNX Vinyl plant assistant, planned in a related B4L study, assumes a self-hosted Typhoon model (see the brief); we recommend keeping a fallback based on OpenThai or ThaiLLM ready until the Typhoon terms are cleared.
4. The market has a gap that a Chiang Mai host could fill. Thai model makers offer only free proofs of concept with rate limits, and none sells a production API from Thailand. A Chiang Mai host that serves these models under a Thai contract, with data kept on the investor's own hardware, would offer something they do not, though Thai GPU clouds such as Float16 and Thai AI Cloud (Bangkok) compete for the same customers.
5. Evaluate models on the target workloads. The published scores are self-reported and inconsistent. We recommend that the investor test the shortlist on the Thai workloads it intends to serve (for example manuals, customer chat and forms) before choosing, bearing in mind that publishing a comparison against Typhoon may count as "competitive benchmarking" under its terms.

Open questions

• Do the OpenTyphoon Terms bind self-hosted Apache 2.0 weights, and does internal evaluation count as "competitive benchmarking"? Ask SCB 10X and SCB DataX, and a Thai law firm.
• Will Typhoon get a paid production API, and where will it run? The AWS launch "expected" in Q1 2026 has not appeared. Ask SCB 10X.
• How much compute did ThaiLLM's continued pretraining use (GPUs, GPU-hours, duration), and would BDI welcome a commercial host serving ThaiLLM models? Ask BDI (the ThaiLLM program office) and ThaiSC at NECTEC.
• How many H100s trained OpenThai 2.0 (3 or 8, since the same page states both), and what does iApp charge for its API? Ask iApp Technology and AIEAT.
• What do the paid tiers of AI for Thai cost? The site shows 10 calls a day for trials, while a third-party directory lists 1,000 a day free (a search result only, not verified). Ask NECTEC's AI for Thai team.
• Can SEA-LION's limit be raised? Its API page invites requests for a higher rate limit, while its FAQ says the free tier cannot be raised. Which cloud partners serve SEA-LION in Thailand? Ask AI Singapore.
• How good are DeepSeek V3.x and Gemma 3 at Thai? We found no primary source; they should be tested on the target workloads.
• Is there an independent Thai leaderboard that compares all these families on one test setup? We fetched none dated 2026; ask VISTEC (WangchanX) and AI Singapore (SEA-HELM).