ThakiCloud/Qwen3.8-27B-Human-KO

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 1, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ThakiCloud/Qwen3.8-27B-Human-KO is a 27 billion parameter Qwen3.8 model fine-tuned for Korean conversational use, featuring a 32768 token context length. It is specifically adjusted to produce short, natural Korean prose and suppresses the generation of Chinese and Japanese tokens. This model excels in human-like conversational style and maintains high performance in coding tasks, making it suitable for Korean-centric applications requiring refined linguistic output.

Loading preview...

ThakiCloud/Qwen3.8-27B-Human-KO Overview

This model is a full-weight checkpoint of Qwen3.8-27B, specifically adapted for Korean conversational use. It has undergone significant adjustments in two key areas to enhance its performance and suitability for Korean language tasks:

Key Differentiators & Capabilities

  • Refined Korean Writing Style: The model was trained with a custom synthetic corpus to merge a specific writing style into its weights. This results in a default output of short, flowing Korean prose, with free generation bullet ratios reduced from 97.5% to 2.0% and average length shifting from 1,326 to 220 characters.
  • CJK Output Suppression: It actively blocks the output pathways for Chinese and Japanese character tokens at the weight level, ensuring cleaner Korean output. Single Chinese characters used in Korean Hanja notation (e.g., 개항(開港)) are preserved. This suppression reduces CJK contamination from 2.55% to 0.33% (with 0.24% residual errors, mostly preserved Hanja).
  • Human-like Conversational Performance: Benchmarks show a significant improvement in human-likeness, with a 94.9% win rate in pairwise human judgment against the base model.
  • Coding Proficiency: Despite language-specific adjustments, the model shows a +3.03 percentage point improvement on HumanEval (n=58) compared to its base, indicating strong coding capabilities are maintained.

Considerations for Use

  • Default Short Responses: The model has a natural tendency for shorter answers. For longer, structured documents, users should specify length and format in their prompts.
  • Language Focus: While optimized for Korean conversation, its performance in English or mathematical tasks has not been extensively re-measured with sufficient samples.
  • Safety Alignment: The model inherits the safety alignment of the base Qwen3.8-27B model, without additional safety training.

This model is ideal for applications requiring high-quality, natural-sounding Korean conversational output with strong coding support, while actively mitigating unwanted CJK character generation.

Finetunes
2 models