ThakiCloud/Qwen3.8-27B-Human-KO-MerchantDesk-v0.1

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ThakiCloud/Qwen3.8-27B-Human-KO-MerchantDesk-v0.1 is a 27 billion parameter Qwen3.8-based model developed by ThakiCloud, specifically fine-tuned for Korean small business owners. It acts as a merchant desk AI, designed to handle customer inquiries for local restaurants by interpreting intent, managing slots, and deciding on appropriate actions. This model excels at policy-based decision-making for customer service, significantly improving resolution rates and intent accuracy compared to general-purpose models, with a context length of 32768 tokens.

Loading preview...

Overview

ThakiCloud/Qwen3.8-27B-Human-KO-MerchantDesk-v0.1 is a specialized 27 billion parameter model built upon ThakiCloud's Korean language model, ThakiCloud/Qwen3.8-27B-Human-KO. It is specifically designed to serve as an AI-powered customer service desk for small business owners in South Korea, particularly for local restaurants. The model's core function is to interpret customer inquiries, identify their intent (from 25 types), extract relevant slots, and determine the next action: direct answering, asking for clarification, using a tool for lookup, or escalating to the owner. It does not memorize factual information like prices or menus but retrieves it from provided store profiles.

Key Capabilities & Performance

This model demonstrates significant improvements over its base model and general-purpose policy models in its target domain:

  • Enhanced Resolution Rate: Achieves a 70.8% resolution rate on IID (Independent and Identically Distributed) datasets, a 10.0%p increase over the base model, and 55.7% on Hard datasets, an 8.3%p increase.
  • High Intent Accuracy: Boasts an 86.6% intent accuracy on IID datasets, a 13.0%p improvement.
  • Reduced Over-escalation: Significantly decreases unnecessary handoffs to the owner, with only 4.2% over-escalation on IID datasets, down from 10.4%.
  • Fact-Grounded Responses: When combined with a compiler, it produces zero instances of unfounded facts in its responses, ensuring reliability.
  • Robust Policy Learning: Benchmarks show that general-purpose policy learning is insufficient for this specific task, highlighting the necessity and effectiveness of this model's dedicated training.

Usage & Limitations

Users should provide a system prompt (system_prompt_v1.2.txt) and customer messages with store profiles. The model outputs a JSON object detailing intent, slots, and the recommended action. It is compatible with Qwen3.8-compatible serving frameworks like vLLM.

Important Note: The model's direct answer output can sometimes contain unfounded facts if used without a compiler. It is strongly recommended to use the model's intent and slots outputs to drive templated responses or rule engines, ensuring factual accuracy by validating against store profiles. This model is currently limited to the restaurant industry and was trained using the AI Hub's "Small Business Customer Order Q&A Text" dataset.

Finetunes
3 models