iapp/ openthai-r1-32b:latest

11 6 days ago

Thai reasoning model that shows its step-by-step thinking — beats DeepSeek R1 70B and Typhoon R1 70B on Thai benchmarks (avg 71.58 vs 63.31/65.42) at half their size. ~24 GB RAM. 

ollama run iapp/openthai-r1-32b

Details

6 days ago

0a038a56ad52 · 20GB

model
qwen2
·
32.8B
·
Q4_K_M
template
{{ if .System }}{{ .System }}{{ end }}{{ if .Prompt }}<｜User｜>{{ .Prompt }}{{ end }}<｜Assistan
license
Available for Research and Commercial use. Licence metadata on the source repo is other/qwen. See ht
params
{ "num_ctx": 32768, "num_predict": 16384, "stop": [ "<｜User｜>", "<

Readme

OpenThai

OpenThaiGPT R1 32B — GGUF

ollama run iapp/openthai-r1-32b

Official GGUF quantizations of openthaigpt/openthaigpt-r1-32b-instruct

Website · Discord · Paper

A Thai reasoning model that shows its work — and beats 70B models at less than half their size. These are the official quantizations, built from the source weights by the OpenThai team.

Quants

File Quant Size Notes
openthaigpt-r1-32b-instruct.Q4_K_M.gguf Q4_K_M ~20 GB Recommended. Best size/quality tradeoff; fits a 24 GB GPU.
openthaigpt-r1-32b-instruct.Q5_K_M.gguf Q5_K_M ~23 GB Higher quality, needs a little more headroom.
openthaigpt-r1-32b-instruct.Q8_0.gguf Q8_0 ~35 GB Near-lossless. For quality-sensitive evaluation.

Usage

llama.cpp

llama-cli -m openthaigpt-r1-32b-instruct.Q4_K_M.gguf \
  -p "อธิบายความแตกต่างระหว่างสัญญาซื้อขายกับสัญญาเช่าซื้อ" \
  -n 4096 --temp 0.6 --top-p 0.95

Ollama

ollama run hf.co/openthaigpt/openthaigpt-r1-32b-instruct-GGUF:Q4_K_M

⚠️ Chat template

This model uses the DeepSeek-R1 template (<｜User｜> / <｜Assistant｜>), not ChatML. Serving it with ChatML produces degraded output that looks superficially correct. The template is embedded in these GGUF files, so llama.cpp and Ollama pick it up automatically — but if you build your own prompt string, use the DeepSeek format.

Recommended sampling: temperature=0.6, top_p=0.95, top_k=40. The upstream card recommends no system prompt for this model.

Benchmarks

Benchmark OpenThaiGPT R1 32B DeepSeek R1 70B Typhoon R1 70B
AIME24-TH 56.67 33.33 53.33
AIME24 63.36 53.33 53.33
MATH500-TH 83.80 75.40 81.00
LiveCodeBench-TH 62.16 53.15 47.75
LiveCodeBench 69.67 64.97 54.79
OpenThaiEval 76.05 74.17 77.59
Average 71.58 63.31 65.42

Quantization costs some accuracy; these figures are from the full-precision model.

Citation

@misc{yuenyong2025openthaigpt16r1thaicentric,
      title={OpenThaiGPT 1.6 and R1: Thai-Centric Open Source and Reasoning Large Language Models},
      author={Sumeth Yuenyong and Thodsaporn Chay-intr and Kobkrit Viriyayudhakorn},
      year={2025},
      eprint={2504.01789},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.01789}
}

OpenThai (formerly OpenThaiGPT) — free, open-weight Thai large language models from AIEAT and iApp Technology. With thanks to the community members who published unofficial GGUF conversions before these existed.