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-based language model, fine-tuned for Korean conversation. It significantly shifts the output style from bulleted lists to short, flowing Korean prose, reducing bullet-list generation from 97.5% to 2.0%. This model also features CJK output suppression, blocking Chinese-word and kana tokens while preserving single-character Hanja for Korean glosses. It is optimized for human-like Korean conversational responses and maintains a 32768 token context length.

Loading preview...

Overview

ThakiCloud/Qwen3.8-27B-Human-KO is a 27 billion parameter model derived from Qwen3.8-27B, specifically adapted for Korean conversation. Its primary focus is on refining the output style and ensuring clean Korean generation by suppressing unwanted CJK characters.

Key Differentiators & Capabilities

  • Korean Style Alignment: The model has been trained on an in-house synthetic corpus to produce short, flowing Korean prose, drastically reducing the rate of bullet-list generation from 97.5% to 2.0% and median response length from 1,326 to 220 characters.
  • CJK Output Suppression: It actively blocks the generation of Chinese-word and kana tokens at the weight level, ensuring a cleaner Korean output. Single-character Hanja used for Korean glosses are preserved.
  • Improved Human-likeness: Achieves a 94.9% win rate in pairwise human-likeness evaluations by an LLM judge, compared to the base model's 1.1%.
  • Code Performance: Shows a slight improvement in HumanEval scores (+3.03 percentage points).

Important Considerations

  • Output Tendency: The model defaults to short answers. For longer, structured documents, explicit length and format instructions are necessary in the prompt.
  • Unmeasured Axes: English and mathematical performance have not been extensively measured with sufficient samples.
  • Safety Alignment: Inherits the safety alignment of the base Qwen3.8-27B model without separate safety training.
  • Hanja Reading: While it can read Hanja, it cannot generate Chinese-word tokens.

Research

The style alignment behind this model is further detailed in the paper "Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model" (arXiv:2609.11291).