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

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ThakiCloud/Qwen3.8-27B-Human-KO-Finance is a 27 billion parameter language model based on the Qwen3.8 architecture, specifically fine-tuned for Korean securities compliance advisory. Developed by ThakiCloud, this model is designed to observe mandatory disclosures and the suitability principle in financial client conversations. It is uniquely engineered with a mandatory gate mechanism to ensure 100% disclosure coverage and zero factual errors in its domain. This model is best suited for applications requiring strict adherence to Korean financial regulations in advisory contexts.

Loading preview...

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

This model is a 27 billion parameter variant of the Qwen3.8 architecture, developed by ThakiCloud, specifically aligned for Korean securities compliance advisory. Its primary function is to ensure adherence to mandatory disclosures and the suitability principle in client interactions without explicit prompting.

Key Differentiators and Capabilities

  • Mandatory Gate Mechanism: A unique feature requiring a deterministic verifier block (gates/) to be attached. This gate is crucial for achieving 100% disclosure coverage and eliminating factual errors, which probabilistic generation alone cannot guarantee.
  • High Compliance: Achieves a suitability behavior pass rate of 0.818 and 0 factual errors (out of 200) when deployed with its gate, significantly outperforming baselines.
  • Regulatory Alignment: Evaluation criteria are based on the Korea Financial Investment Association's Standard Investment Recommendation Guidelines and Article 19 of the Financial Consumer Protection Act.
  • Specialized Training: Fine-tuned from ThakiCloud/Qwen3.8-27B-Human-KO using synthetic data generated by ThakiCloud's own serving model.

Important Limitations

  • The model must be used with its accompanying gate; weights alone do not guarantee compliance.
  • Suitability judgments are based on a limited n=33 items, making absolute values less reliable than paired comparisons.
  • The situation classifier has an accuracy of 0.885, which is mitigated by using a union of input classification and output-based judging for disclosures.
  • Legal review is required before deployment in actual regulatory environments, as the gate only backfills shortfalls and does not prevent all missing disclosures by itself.

This model represents the securities-compliance arm of a study detailed in arXiv:2609.11291, focusing on emission policy results in compliance and safety stages.

Finetunes
1 models