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

TEXT GENERATIONConcurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:32kPublished:Jul 17, 2026License:nvidia-open-model-agreementArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The iapp/openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b is a 30 billion parameter Mixture-of-Experts (MoE) model developed by the OpenThai team (AIEAT / iApp Technology) based on NVIDIA's Nemotron-3-Nano-Omni-30B-A3B-Reasoning. This model is specifically fine-tuned for Thai legal knowledge, excelling at verifiable, section-level citation from Thai statutes and court rulings. It is designed for self-hostable legal applications, offering strong performance in closed-book recall and legal essay generation in Thai.

Loading preview...

OpenThai 2.0 Legal ThaiLLM: Specialized for Thai Legal Applications

The openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b is an open-weight, 30 billion parameter Mixture-of-Experts (MoE) model, with approximately 3 billion active parameters per token. Developed by the OpenThai team (AIEAT / iApp Technology) on NVIDIA's Nemotron-3-Nano-Omni-30B-A3B-Reasoning base, this model is uniquely optimized for Thai legal knowledge and citation.

Key Capabilities

  • Legal Citation Accuracy: Cites exact law names and section numbers in a JSON format, crucial for RAG and legal drafting tools.
  • Extensive Legal Knowledge: Demonstrates strong closed-book recall of Thai statutes, outperforming Qwen3.6-35B on Civil and Commercial Code recall by ~4x.
  • Legal Essay Generation: Excels in generating legal essays, surpassing Qwen3.6-35B across four axes: citations, holding, coverage, and fluency.
  • Grounded Training: Trained on Thai statutes and court rulings, ensuring answers are tied to verifiable source sections.
  • Efficient Deployment: Runs on a single 80 GB GPU (or 24 GB with NVFP4 build) and supports vLLM and NVIDIA NIM for OpenAI-compatible API serving.

Training Methodology

The model underwent a three-stage training process using the NVIDIA NeMo stack:

  1. Continued Pretraining (CPT): Ingested Thai statutes and legal rulings to embed legal knowledge within its weights.
  2. Supervised Fine-tuning (SFT): Taught grounded answering, ensuring outputs adhere to a fixed JSON contract citing only provided sections.
  3. Reinforcement Learning (GRPO): Refined precision using the official benchmark's citation F1 score as a reward, penalizing incorrect or missing citations.

Good for

  • Legal Research: Answering questions based on Thai statutory law and court rulings, both with and without retrieval.
  • Automated Legal Drafting: Generating legal text with precise, verifiable citations.
  • Legal Tech Pipelines: Integrating structured legal outputs into applications requiring high citation accuracy.
  • Self-Hosted Solutions: Deploying a powerful Thai legal LLM on local infrastructure with manageable hardware requirements.

Finetunes
3 models