xidrug/qwen3.5-9b-kh-trade-logistics-merged-v1

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 22, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The xidrug/qwen3.5-9b-kh-trade-logistics-merged-v1 is a 9.4 billion parameter language model based on Qwen/Qwen3.5-9B, fine-tuned for Chinese trade and logistics question answering. It was developed by xidrug by merging a QLoRA adapter, specifically optimized for domain-specific inquiries. This model is designed to function as a component in Retrieval Augmented Generation (RAG) systems for trade and logistics, supporting a 32K context length.

Loading preview...

Overview

This model, xidrug/qwen3.5-9b-kh-trade-logistics-merged-v1, is a specialized 9.4 billion parameter language model built upon the Qwen/Qwen3.5-9B base. It has been fine-tuned using a 4-bit NF4 QLoRA SFT method, specifically for question answering within the Chinese trade and logistics domain. The released model is a merged bf16 full weight version, eliminating the need for separate QLoRA adapter loading.

Key Capabilities

  • Domain-Specific QA: Optimized for answering questions related to Chinese trade and logistics.
  • RAG Component: Designed to integrate effectively as an auxiliary component in Retrieval Augmented Generation (RAG) workflows.
  • Training Data: Trained on a substantial dataset of 99,450 samples, with a final training loss of 0.464494.
  • Performance: Achieved an 83.46% character-level F1 score and an 82.81% ROUGE-L F1 score on a stratified independent test set of 86 samples.
  • Context Length: Supports a context length of 32,768 tokens.

Good For

  • Trade and Logistics QA Systems: Ideal for applications requiring accurate responses to queries within the Chinese trade and logistics sector.
  • RAG Implementations: Suitable for developers building RAG systems where domain-specific knowledge retrieval and synthesis are crucial.

Limitations

  • Dynamic Information: Not recommended for real-time dynamic information (e.g., policies, rates, prices, port status) without external, up-to-date knowledge bases and human review.
  • JSON Extraction: The model shows a 0.00% legitimate JSON ratio in current independent tests, indicating that structured JSON output requires additional training, validation, or constrained decoding.

Finetunes
1 models