GGUF conversion and quantization of iapp/openthai2.0-legal-thaillm-nemotron-3-nano-30b-a3b, an open-weight Thai legal LLM built on a NVIDIA Nemotron-3-Nano-Omni-30B-A3B-Reasoning base (Mamba2-Transformer hybrid MoE, 30B total / ~3B active parameters per token). The base model exposes exact statute and มาตรา (section) citations in a structured JSON contract, tuned for both closed-book recall and retrieval-augmented (RAG) legal question answering.

This repo provides quantized .gguf files for use with llama.cpp and compatible runtimes (Ollama, LM Studio, Jan, koboldcpp, etc). For the model card, benchmark results, training details, and usage examples, see the original repository.

Note the 30B MoE with ~3B active parameters per token means these files are larger than a dense 30B model but should run at speeds closer to a ~3B dense model, VRAM/RAM permitting.

Converted from the original safetensors with llama.cpp's convert_hf_to_gguf.py, then quantized with llama-quantize:

or as an OpenAI-compatible server:

The model supports a served context window of 32,768 tokens. See the original model card for the chat template, recommended generation settings (temperature/top_p per use case), and JSON citation-answer prompting format — these apply unchanged to the GGUF builds.

Distributed under the same NVIDIA Open Model Agreement as the base model and the original OpenThai 2.0 Legal release. This is a format conversion and quantization only — no weights were retrained or modified beyond quantization. All credit for the underlying model goes to the OpenThai team (AIEAT / iApp Technology) and NVIDIA.

As with the source model: outputs are decision support, not legal advice, and citations should be verified against current law by a qualified professional before being relied upon.