KBTG-Labs/ThaiLLM-8B-Instruct

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Dec 17, 2025License:apache-2.0Architecture:Transformer Open Weights Warm

KBTG-Labs/ThaiLLM-8B-Instruct is an 8 billion parameter instruction-following language model developed by KBTG-Labs, built by merging ThaiLLM-8B and Qwen3-8B using mergekit. It features enhanced Thai language understanding and supports switching between thinking and non-thinking modes, similar to Qwen3-8B. This model is primarily designed for applications requiring strong instruction-following capabilities and improved performance in Thai language contexts, demonstrating competitive results on Thai-specific benchmarks.

Loading preview...

ThaiLLM-8B-Instruct Overview

ThaiLLM-8B-Instruct is an 8 billion parameter language model developed by KBTG-Labs, specifically designed to enhance instruction-following capabilities. It is constructed using mergekit, integrating the strengths of ThaiLLM-8B for robust Thai language understanding and Qwen3-8B for its advanced features.

Key Capabilities

  • Enhanced Thai Language Understanding: Leverages ThaiLLM-8B to provide superior comprehension and generation in Thai.
  • Flexible Thinking Modes: Supports dynamic switching between 'thinking' and 'non-thinking' modes, a feature inherited from Qwen3-8B, allowing for adaptable reasoning processes.
  • Improved Instruction Following: Fine-tuned to better adhere to user instructions, making it suitable for a variety of task-oriented applications.

Good For

  • Thai-centric Applications: Ideal for use cases requiring high proficiency in the Thai language, such as chatbots, content generation, and summarization for Thai users.
  • Instruction-Following Tasks: Excels in scenarios where precise adherence to prompts and instructions is critical.
  • Research and Development: Provides a strong base for further fine-tuning or experimentation, particularly in multilingual or Thai-specific NLP research.

Benchmarking results indicate that ThaiLLM-8B-Instruct shows competitive performance on various exams, including M3 Exam, M6 Exam, Flare CFA, and IC, often outperforming Qwen3-8B in non-thinking mode and showing strong results in thinking mode. For a detailed analysis, refer to the Technical Report.

Finetunes
1 models