Run with vLLM
vllm serve iapp/OpenThai2.1-Nemotron-3.5-Bolt-E4B-NVFP4 \ --served-model-name openthai-e4b \ --max-model-len 16384 \ --enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4 \ --limit-mm-per-prompt '{"image":0,"video":0,"audio":0}'
Thinking is off by default; chat_template_kwargs={"enable_thinking": true} turns it on.
We tested it on one H100 with vLLM v0.31.0. H100 has no FP4 hardware, so there vLLM runs the NVFP4 layers as 4-bit
weights with 16-bit activations (its Marlin kernel). Blackwell GPUs also quantize the activations to 4 bits; we have
not tested that path yet.
Tests
Table with columns: Hardware, Test, Result| Hardware | Test | Result |
|---|
| H100 (vLLM v0.31.0) | load and "สวัสดีครับ คุณคือใคร ตอบสั้น ๆ" | loads; "ผมชื่อ OpenThai 2.1 Bolt E4B ผู้ช่วย AI ภาษาไทยครับ" |
| H100 (vLLM v0.31.0) | ThaiExam, OpenThaiEval, IFEval-TH (thinking off, greedy) against the BF16 model | see below |
| H100 (vLLM v0.31.0) | the command above on a fresh download of this repository (with --gpu-memory-utilization 0.25 added, because the GPU was shared): a self-introduction, a short answer, thinking on, a weather tool call with its result sent back, and a second turn | all passed; every downloaded file is identical to the build we tested |
| DGX Spark | load and a Thai self-introduction | not tested yet |
Table with columns: NVFP4, BF16, Δ, SE | NVFP4 | BF16 | Δ | SE |
|---|
| ThaiExam (565) | 66.1 | 67.3 | −1.21 | 1.31 |
| OpenThaiEval (1,232) | 67.1 | 68.0 | −0.89 | 0.78 |
| IFEval-TH (215) | 79.1 | 79.5 | −0.47 | 0.81 |
None of the three drops is larger than twice its standard error, but all three point the same way. If memory allows,
the BF16 model is the reference.
We built this version twice. With ModelOpt's default preset, which takes each block's scale from its largest weight,
ThaiExam was 3.4 points below BF16 (SE 1.5). The build in this repository uses the error-minimising scales instead;
it was the only other preset we scored. A third, activation-aware preset (AWQ) wrote a format that vLLM v0.31.0 does
not load.
About the model
OpenThai 2.1 Bolt E4B คือโมเดลภาษาขนาดเล็กแบบเปิดน้ำหนัก สำหรับใช้งานภาษาไทยบนเครื่องของผู้ใช้เอง
ทีม OpenThai (บริษัท ไอแอพพ์เทคโนโลยี จำกัด ร่วมกับสมาคมผู้ประกอบการปัญญาประดิษฐ์ประเทศไทย AIEAT) ฝึกเพิ่มจาก
NVIDIA Nemotron 3.5 Bolt E4B (8B พารามิเตอร์ ใช้งานจริง 4B) ด้วยข้อมูลความรู้ไทย ข้อมูลตัวตนของ OpenThai
คำตอบภาษาไทยที่โมเดลฐานเขียนเอง และบทสนทนาการเรียกใช้เครื่องมือ (tool calling) โมเดลนี้เป็นโมเดลเริ่มต้นของ ChindaBot และตอบแบบไม่คิด (thinking off)
เป็นค่าเริ่มต้น
OpenThai 2.1 Bolt E4B is a small open-weights model for Thai on your own computer. The OpenThai team (iApp
Technology with AIEAT) fine-tuned it from NVIDIA Nemotron 3.5 Bolt E4B (8B parameters, 4B effective) on Thai
knowledge, OpenThai identity, Thai answers written by the base model itself, and tool-calling conversations. It is the starter model of ChindaBot
and answers with thinking off by default.
Results of the BF16 model
Thinking off (greedy, 4k new tokens). All rows were run with the same harness and settings.
Table with columns: Model, ThaiExam, OpenThaiEval, IFEval-TH, MT-Bench-Thai, MATH-500-TH, Code-switching, IFEval (en), MMLU-Redux (en), Thai answer rate| Model | ThaiExam | OpenThaiEval | IFEval-TH | MT-Bench-Thai | MATH-500-TH | Code-switching | IFEval (en) | MMLU-Redux (en) | Thai answer rate |
|---|
| OpenThai 2.1 Bolt E4B | 67.3 | 68.0 | 79.5 | 6.05 | 75.4 | 98.5 | 86.7 | 83.8 | 98.4 |
| Nemotron 3.5 Bolt E4B (base) | 67.3 | 67.7 | 80.0 | 6.12 | 75.8 | 99.5 | 86.5 | 83.4 | 98.4 |
| OpenThai 2.1 Bolt 74B-A9B | 63.4 | 72.5 | 72.1 | 5.38 | 79.2 | 98.0 | 82.8 | 89.5 | 96.2 |
- ThaiExam: Thai university-entrance and professional exams (O-NET, A-Level, TGAT, TPAT1, IC), 565 questions.
- OpenThaiEval: 1,232 Thai multiple-choice questions.
- IFEval-TH: Thai instructions with checkable format rules, 215 prompts.
- MT-Bench-Thai: 91 two-turn Thai conversations, scored 1-10 by OpenThai 2.0 as judge (a model from our own family).
- Thai answer rate: share of MT-Bench-Thai answers written in Thai.
Release gate
We wrote these rules before scoring any trained model. A model is released only if it passes all four
(Δ = trained minus base, SE = paired standard error):
- Thai knowledge: the mean Δ of ThaiExam and OpenThaiEval is above 0, and neither is below −2·SE.
- Other Thai skills: IFEval-TH, code-switching and MATH-500-TH each have Δ ≥ −2·SE.
- English: IFEval and MMLU-Redux each have Δ ≥ −2·SE.
- Thai answers: the Thai answer rate is at least 95%.
Result: passed. None of the differences from the base is significant; the largest drop is code-switching
(two prompts of 200).
Table with columns: Δ, SE, Rule | Δ | SE | Rule |
|---|
| ThaiExam (565) | +0.01 | 0.85 | 1 |
| OpenThaiEval (1,232) | +0.32 | 0.38 | 1 |
| IFEval-TH (215) | −0.47 | 1.23 | 2 |
| Code-switching (200) | −1.00 | 1.00 | 2 |
| MATH-500-TH (500) | −0.40 | 1.10 | 2 |
| IFEval (541) | +0.18 | 0.85 | 3 |
| MMLU-Redux (5,700) | +0.46 | 0.32 | 3 |
| Thai answer rate | 98.4% | | 4 |
The Thai-knowledge mean Δ is +0.17, which passes rule 1 but is well within noise: we do not claim that the model
knows more Thai than the base. What the training adds is the OpenThai identity; the tool-calling conversations keep
tool use at the base model's level (table below).
MT-Bench-Thai, which is not part of the gate, is 0.07 lower than the base (paired SE 0.13 over 91 conversations,
one judging run). Also outside the gate, with thinking off: HumanEval 90.2 against the base model's 89.0,
HumanEval-TH 75.6 against 72.6 and HellaSwag-TH 53.3 against 52.8; none of these differences is significant.
English tool calling (BFCL v3: accuracy over all 3,841 items of 14 categories, thinking off) is 79.8, the same as
the base model's.
How this version was chosen. Each version was judged by a rule written before it was trained.
- The first run that passed the gate had no tool conversations in its data and often ended its turn after a tool
result without answering (9 of the 16 cases below, llama.cpp). It was not released.
- The second added tool conversations and was the first public release (11 November 2026). It was 2.2 points below
the base on MATH-500-TH.
- The third used math and instruction-following answers only if a program verified them, and also taught the model
to write a short reason before a multiple-choice answer. ThaiExam fell by 4.8 points; it failed the gate and was
not released.
- The fourth left out those multiple-choice rows. The math gap was gone, but HumanEval fell to 84.1 against the base
model's 89.0, and it was not released.
- This version, the second release (15 November 2026), adds English coding answers that pass unit tests. Its rule:
pass the gate; answer after the tool result in at least 15 of the 16 cases; name OpenThai in at least 115 of 120
identity answers; hold the first release's level on ThaiExam and OpenThaiEval and score higher on the mean of the
five Thai benchmarks; not fall below the first release on MT-Bench-Thai; be no lower than the base by more than
twice the standard error on HumanEval, HumanEval-TH and HellaSwag-TH; and have a Q4_K_M file no worse than the
first release's. It passed all of them.
Table with columns: This model, Base | This model | Base |
|---|
| Identity: answers that name OpenThai, of 120 (8 questions, asked 3 ways, 1 greedy and 4 sampled answers each) | 120 | 0 |
| Identity: the same with 8 questions worded differently from the training data, of 120 | 120 | 0 |
| Tools: right tool on the first call, of 16 cases (Thai and English, 6 tools; greedy) | 16 | 16 |
| Tools: an answer after the tool result, of 16 (llama.cpp b12242, Q4_K_M) | 16 | 16 |
| Tools: an answer after the tool result, of 16 (vLLM, BF16) | 16 | 14 |
Thinking off. Sampled identity answers use temperature 0.7. The base model named another developer in 67 of its 120
identity answers. The tool rows count non-empty answers; we read this model's 16 llama.cpp answers, and each one used
the tool result and answered in the language of the question.
Training data
81,856 conversations, of which 300 were held out for validation. All are thinking off.
Table with columns: Part, Rows, Answers written by| Part | Rows | Answers written by |
|---|
| Thai fact questions and answers (OpenThai 2.0 fact base) | 25,601 | stored answers, unchanged |
| Thai general instructions, grounded reading and summaries (OpenThai 2.0 prompts) | 13,922 | the base model (self-distillation) |
| Instruction following with format rules, 5,251 Thai and 2,249 English (OpenThai 2.0 constraints, NVIDIA IF-Chat v4 prompts); kept only if a program confirms every rule | 7,500 | the base model, verified |
| Thai free recall from laws and documents, short and long answers | 4,495 | stored answers, unchanged |
| Thai number copying from documents | 1,985 | the base model |
| Thai two-turn chat | 1,988 | the base model |
| English chat | 1,965 | the base model |
| Thai mathematics (nvidia/OpenMathInstruct-2 problems, translated); kept only if the final answer is right | 5,577 | the base model, verified |
| English mathematics (nvidia/OpenMathInstruct-2 problems); kept only if the final answer is right | 2,000 | the base model, verified |
| English coding, 1,000 instructions and 1,000 function stubs (nvidia/OpenCodeInstruct problems); kept only if the code passes the problem's unit tests | 2,000 | the base model, verified |
| Thai coding | 1,748 | the base model |
| Thai multiple choice, letter-only answers | 2,313 | stored answers, unchanged |
| OpenThai identity (Thai and English) | 1,762 | authored by the OpenThai team |
| Thai poetry | 1,000 | stored answers, unchanged |
| Royal vocabulary (ราชาศัพท์) | 817 | stored answers, unchanged |
| Tool conversations, 3,744 Thai and 3,439 English (nvidia/Nemotron-SFT-Agentic-v2 prompts and tools; the Thai prompts are the translations used for OpenThai 2.1 Bolt): the next tool call, or the answer after the tool results | 7,183 | the base model |
Why most answers are the base model's own. The E4B base already scores higher on ThaiExam than our 74B model. A
first training run that used answers written by larger OpenThai models lowered its ThaiExam score from 67.3 to 59.9
and IFEval-TH from 80.0 to 72.6, and it was not released. In this version the base model writes the general answers
itself (thinking off, greedy). Only the fact, free-recall, multiple-choice, poetry and royal-vocabulary rows keep their
stored answers, and the identity rows teach the OpenThai name.
Verified answers. For mathematics, instruction following with format rules and English coding, the base model
answered each prompt two or three times with sampling, and an answer was kept only if a program verified it: the
final answer is right, every format rule holds, or the code passes the problem's unit tests (84% of the coding
problems had a passing answer). No benchmark question is used, and coding problems that share a function name or a
50-character passage with a HumanEval prompt were left out.
Base-model answers were kept only if they finished, contained no chat-template markers or foreign scripts, did not
name another AI model or company, and, for Thai prompts, were mostly Thai. 98% of them passed. Tool calls were also
kept only if they named one of the offered tools and their arguments were a JSON object; 96% of the tool-conversation
answers passed.
Limitations
- A 4B-effective model knows less than large models. Check facts that matter, especially names, dates, numbers
and legal citations.
- Images were checked, not benchmarked; audio was not tested. The vision and audio encoders are the base model's,
unchanged. In ChindaBot's setup (llama.cpp b12242, Q4_K_M with the image projector) the model named three coloured
shapes and read the total and items of a printed English receipt correctly.
- With thinking on, scores are slightly below the base model's. Training used thinking-off conversations only.
With thinking on (temperature 1.0, top-p 0.95, up to 30k new tokens, one run each), this model scores IFEval-TH
80.0 against the base model's 82.8, MATH-500-TH 77.6 against 78.2 and ThaiExam 70.2 against 71.4. Each gap is less
than twice its standard error, but all three point the same way.
- MT-Bench-Thai is judged by OpenThai 2.0, a model from our own family.
- The model can produce incorrect, biased or unsafe content. It inherits the base model's safety tuning, and we added
no safety-specific training.
Citation
@misc{openthai21bolte4b2026, title = {OpenThai 2.1 Bolt E4B}, author = {Viriyayudhakorn, Kobkrit and {OpenThai Team}}, year = {2026}, note = {iApp Technology and AIEAT}}
Built on NVIDIA Nemotron 3.5 Bolt E4B (OpenMDW-1.1), which NVIDIA post-trained from Google Gemma 4 E4B-it; use of the
Gemma 4 model is governed by the Apache License, Version 2.0. Training data attributions: OpenThai 2.0 training data
(iApp/AIEAT), nvidia/Nemotron-SFT-Instruction-Following-Chat-v4 prompts (OpenMDW-1.1), nvidia/Nemotron-SFT-Agentic-v2
prompts and tools (CC-BY-4.0/Apache-2.0/MIT), nvidia/OpenMathInstruct-2 prompts (CC-BY-4.0), nvidia/OpenCodeInstruct
problems (CC-BY-4.0). Compute: Siam AI Corporation.
The license texts and the NOTICE are in this repository.