⚖️ Introducing OpenThai 2.0 Legal

Announced 24 July 2026

The Thai legal AI that knows the law by heart — built by Thais, for Thailand.

The OpenThai project is releasing OpenThai 2.0 Legal, an open-weight large language model purpose-built for Thai law, with its full weights freely available on Hugging Face.

It is designed for lawyers, in-house counsel, compliance teams, legal-tech builders and public agencies that need reliable, source-grounded answers on Thai statutes and case law.

Try it live

Talk to the real model in your browser — nothing to install.

Live demo + full documentationiapp.co.th/openmodels/openthai2p0-legal

That page hosts an interactive demo covering all three modes (closed-book / citation-RAG / legal essay), plus a hosted API that is free until 24 August 2026 (a free iApp API key is required).

Download the model

OpenThai 2.0 Legal 30B-A3B — Hugging Facehuggingface.co

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

🌟 Highlights

• Open weights, self-hostable — a 30B Mixture-of-Experts model activating only ~3B parameters per token, so it runs on a single GPU from 24 GB of VRAM in NVFP4 (4-bit).
• Cites the law, not vibes — returns the exact law name and section (มาตรา) in a fixed JSON contract, ready for RAG pipelines and drafting tools.
• Knows the statute book from memory — asked with no statute in the prompt, it recalls the right section: roughly 4× the closed-book recall of the much larger Qwen3.6-35B on the Civil and Commercial Code.
• Writes like a lawyer — ahead of Qwen3.6-35B on all four legal-essay axes.
• Grounded in real Thai legal sources — trained on published Thai statutes and Supreme Court rulings.

📊 Benchmark results

Evaluation	OpenThai 2.0 Legal	Qwen3.6-35B	Nemotron-3-30B base
Closed-book: knows the law from memory
Civil & Commercial Code (n=3,729)	0.07	0.02	0.001
Tax — Revenue Code (n=50)	0.40	0.36	0.31
Open-book: uses provided law (RAG)
Civil & Commercial, cite from context (n=3,729)	0.99	0.99	0.98
Tax echo (n=50)	0.84	0.84	0.64
Tax selection among distractors (n=50)	0.69	0.64	0.45
Legal essays (Thai Supreme Court cases)
Correct citations (n=72)	0.25	0.09	0.02
Reaches the right holding (n=72)	0.57	0.50	0.31
Covers the key points (n=72)	0.60	0.55	0.27
Writing quality (n=72)	0.46	0.43	0.13

Scores run 0 to 1, higher is better. Citation rows use NitiBench's citation-F1 scorer; holding, coverage and fluency are judged by Gemini 3.1 Flash Lite. Every model was re-run end to end under one identical protocol — single pass, thinking off, citations checked by code.

ℹ️ Scope of the claim

This is the strongest closed-book Thai-law recall measured in a self-hostable model. A closed cloud API still scores higher.

🧾 Model details

Model name	openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b
Base model	NVIDIA Nemotron-3-Nano-Omni-30B-A3B-Reasoning (text core extracted; Mamba2-Transformer hybrid MoE)
Parameters	30B total, ~3B active per token
Context length	256k (262,144 tokens), prompt plus completion
Languages	Thai (primary), English (reasoning)
Precision	bfloat16 safetensors, 17 shards · NVFP4 for single-GPU 24 GB serving
Training stack	NVIDIA NeMo (Megatron-Bridge for CPT/SFT, NeMo-RL for GRPO)
Serving	vLLM (verified) and NVIDIA NIM (OpenAI-compatible)
License	NVIDIA Open Model Agreement

🔬 How it was trained

Three stages on the NVIDIA NeMo stack:

1. CPT (continued pretraining) — drills the full text of every section in both directions so the law lives in the weights. 360,985 drills, ~800M Thai-law tokens.
2. SFT (supervised fine-tuning) — teaches grounded answering under a fixed JSON contract. 16,436 exam answers and เนติบัณฑิต essays.
3. GRPO (reinforcement learning) — reward is the benchmark's own citation F1, penalizing wrong or missing citations. 8,568 graded questions, ~68,500 drafts.

🚀 Getting started

Serve with vLLM on 2 GPUs:

vllm serve iapp/openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b \
  --tensor-parallel-size 2 --trust-remote-code \
  --max-model-len 32768 --enforce-eager

For the full 256k window, set --max-model-len 262144. No extra flag is needed — 262,144 is the model's own max_position_embeddings — but budget more VRAM for KV cache (~1.5 GiB per full-length sequence).

Or call the hosted API (free until 24 Aug 2026 with a free iApp API key):

from openai import OpenAI
client = OpenAI(base_url="https://api.iapp.co.th/v3/llm/openthai2p0-legal",
                api_key="YOUR_IAPP_API_KEY")

r = client.chat.completions.create(
    model="openthai2.0-legal",
    messages=[{"role": "user", "content": "ลักทรัพย์ในเวลากลางคืน ผิดมาตราใด"}],
    max_tokens=1024,
)
print(r.choices[0].message.content)

🔗 Use it with RAG (strongly recommended)

The model is at its best when the law is in the prompt: open-book citation accuracy is 0.99 versus 0.07–0.40 from pure memory. Always pair it with retrieval for real work.

OpenThaiRAG — Thai-native RAG frameworkopenthai.ai/openthairag

Tutorial: Open WebUI & OpenThaiRAGiapp.co.th

Recommended Thai legal datasets for your retrieval corpus: WangchanX-Legal-ThaiCCL-RAG, NitiBench, and authoritative statute text from the Office of the Council of State.

⚠️ Limitations and responsible use

• Retrieval quality drives results — if the retriever supplies the wrong sections, the answer follows them.
• Near-miss rejection is the hardest task — telling a governing section from a closely related one is the lowest-scoring axis for every model, ours included.
• Scope — trained on Thai statutory law and Revenue Department rulings; niche areas and very recent amendments may be thinner.
• Human oversight required — outputs are decision support, not legal advice. Verify every citation against the current law.

🙏 Credits

Built by the OpenThai team (AIEAT / iApp Technology) on the NVIDIA Nemotron and NeMo stack, together with the ThaiLLM initiative and NVIDIA.

Evaluation uses NitiBench (VISAI-AI, MIT); SFT data built from WangchanX-Legal-ThaiCCL (VISTEC, MIT).