sthaps/openthaigpt-thaillm-8b-instruct-v0.7.2-research-preview-light-uncen

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Apr 14, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The sthaps/openthaigpt-thaillm-8b-instruct-v0.7.2-research-preview-light-uncen is an 8-billion parameter experimental Thai language model, developed by sthaps as a decensored version of the OpenThaiGPT/ThaiLLM model. Based on Qwen3-8B, it is specifically fine-tuned for Thai knowledge Q&A, excelling in areas like Thai history, culture, and government information. This model supports both Thai and English, featuring Chain-of-Thought reasoning and robust understanding through paraphrase augmentation, making it ideal for Thai FAQ and context-specific question-answering tasks.

Loading preview...

Model Overview

This is a decensored version of the OpenThaiGPT-ThaiLLM-8b-instruct-v0.7.2-research-preview model, created by sthaps using the Heretic tool. The original model, developed by OpenThaiGPT in collaboration with ThaiLLM, is an experimental 8-billion parameter Thai language model based on Qwen3-8B.

Key Capabilities

  • Deep Thai Knowledge: Specializes in answering questions about Thai history, culture, government agencies, and general Thai context.
  • Chain-of-Thought Reasoning: Supports step-by-step reasoning using a <think>...</think> format.
  • Paraphrase Robustness: Understands diverse question phrasings due to Paraphrase Augmentation during training.
  • Bilingual Support: Functions in both Thai and English.
  • High IFEval Score: Demonstrates strong instruction following, with IFEval at 87.6% and IFEval-TH at 75.5%.

Performance and Limitations

While optimized for Thai knowledge, the model shows lower performance in competitive mathematics and coding benchmarks compared to general-purpose models. For instance, its AIME24-TH score is 0.0667, significantly lower than Qwen3-8B-Instruct's 0.6667. It is not designed for advanced mathematics or extensive code generation. Users should be aware of potential hallucinations and verify critical information.

Good For

  • Thai FAQ systems.
  • Question answering related to Thai history, culture, and government.
  • Applications requiring robust understanding of Thai-specific contexts.

Finetunes
1 models