iapp/openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b
The iapp/openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b model is an open-weight 30B MoE (Mixture of Experts) language model developed by the OpenThai team (AIEAT / iApp Technology), built on NVIDIA Nemotron-3-Nano-Omni-30B-A3B-Reasoning. It features a 256k token context window, enabled by a Mamba2-Transformer hybrid architecture that significantly reduces KV cache requirements. This model is specifically designed for Thai legal applications, excelling at closed-book recall of Thai statutes, accurate citation generation in JSON format, and producing high-quality legal essays.
Loading preview...
OpenThai 2.0 Legal ThaiLLM: Specialized for Thai Law
openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b is an open-weight 30B Mixture of Experts (MoE) model, with approximately 3B active parameters per token, developed by the OpenThai team (AIEAT / iApp Technology). Built upon the NVIDIA Nemotron-3-Nano-Omni-30B-A3B-Reasoning base, this model is uniquely engineered for Thai legal knowledge and verifiable, section-level citation.
Key Capabilities & Differentiators
- Extended Context Window: Features a 256k (262,144-token) context length, allowing for processing long legal documents and multi-document ruling bundles. This is achieved efficiently through a Mamba2-Transformer hybrid architecture, which drastically reduces KV cache usage compared to dense models.
- Thai Legal Expertise: Demonstrates strong performance in recalling Thai statutes from memory, outperforming Qwen3.6-35B on closed-book Civil and Commercial Code recall by approximately 4x.
- Precise Citation Generation: Capable of citing exact law names and section numbers in a structured JSON format, making it ideal for RAG and legal drafting tools.
- High-Quality Legal Writing: Excels in generating legal essays, surpassing Qwen3.6-35B across four key legal-essay metrics (citations, holding, coverage, fluency).
- Grounded Training: Trained on Thai statutes and court rulings, with answers tied to source sections, ensuring high factual accuracy.
Training Methodology
The model underwent a three-stage training process using the NVIDIA NeMo stack:
- CPT (Continued Pretraining): Ingested Thai statutes and legal rulings to embed legal knowledge directly into the model's weights.
- SFT (Supervised Fine-Tuning): Taught grounded answering, ensuring every response adheres to a fixed JSON contract that cites only provided sections.
- GRPO (Reinforcement Learning): Refined precision using the official benchmark's citation F1 score as a reward, penalizing incorrect or missing citations.
Use Cases
This model is particularly well-suited for applications requiring deep understanding and generation of Thai legal text, including legal research, automated legal drafting, and RAG systems for legal document analysis. It supports both closed-book (pure memory) and open-book (RAG) modes, with robust performance in citing applicable sections from provided contexts.