---

license: other

license_name: nvidia-open-model-agreement

license_link: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-agreement/

language:

• th
• en

base_model: iapp/openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b

pipeline_tag: text-generation

library_name: gguf

tags:

• thai
• legal
• openthai
• gguf
• llama.cpp
• mixture-of-experts

---

<div align="center">

<img src="https://huggingface.co/spaces/openthaigpt/README/resolve/main/openthai-logo-white.png" width="160" alt="OpenThai">

OpenThai 2.0 Legal 30B-A3B — GGUF

Official GGUF quantizations of iapp/openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b

Website · Announcement · Live demo · Discord

</div>

An open-weight Thai legal LLM that recalls Thai statutes and cites the exact law name and

section (มาตรา) as structured JSON. 30B Mixture-of-Experts with only ~3B parameters active

per token — which is why a 30B model runs comfortably on modest hardware.

Use it with retrieval. Open-book citation accuracy is 0.99 versus 0.07–0.40 from pure

memory. Pair it with OpenThaiRAG or your own

retrieval over authoritative statute text.

Quants

| File | Quant | Size | Notes |

|---|---|---|---|

| openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b.Q4_K_M.gguf | Q4_K_M | ~18 GB | Recommended. Fits a 24 GB GPU. |

| openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b.Q5_K_M.gguf | Q5_K_M | ~21 GB | Higher quality. |

| openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b.Q8_0.gguf | Q8_0 | ~32 GB | Near-lossless. |

Usage

Ollama

ollama run hf.co/iapp/openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b-GGUF:Q4_K_M

llama.cpp

llama-cli -m openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b.Q4_K_M.gguf \
  -p "ลักทรัพย์ในเวลากลางคืน ผิดมาตราใด" -n 1024 --temp 0.3

The chat template is embedded in the GGUF. Recommended sampling: temperature=0.3,

top_p=0.9. Architecture is a hybrid Mamba2-Transformer MoE (NVIDIA

Nemotron-3-Nano-30B-A3B base) — use a recent llama.cpp build.

⚠️ Responsible use

Outputs are decision support, not legal advice. Verify every citation against the

current statute text. Near-miss rejection — telling the governing section from a closely

related one — is the hardest task for every model tested, this one included.

Citation

@misc{openthai2026legal,
  title  = {OpenThai 2.0 Legal: An Open-Weight Thai Legal Language Model},
  author = {Viriyayudhakorn, Kobkrit and Yuenyong, Sumeth and Chay-intr, Thodsaporn},
  year   = {2026},
  url    = {https://openthai.aieat.or.th/openthai2p0-legal}
}

---

*OpenThai (formerly OpenThaiGPT) — free, open-weight Thai large language models from AIEAT

and iApp Technology, built here on the NVIDIA Nemotron and NeMo stack. With thanks to the

community members who published unofficial GGUF conversions before these existed.*