nowdoor/Qwen3.8-27B-Inspect-v2

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 19, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

nowdoor/Qwen3.8-27B-Inspect-v2 is a 27 billion parameter experimental language model developed by nowdoor, fine-tuned for reviewing Korean facility safety inspection and precision safety diagnosis reports. This model, a successor to Qwen3.8-27B-Inspect-v01, demonstrates improved accuracy in facility assessment and robust adherence to JSON schema, while maintaining general reasoning performance. It is specifically designed to assist in the structured review and quality control of technical reports within the facility safety domain.

Loading preview...

Overview

nowdoor/Qwen3.8-27B-Inspect-v2 is a 27 billion parameter experimental model, building upon the Qwen3.5 architecture, specifically fine-tuned for the analysis of Korean facility safety inspection and precision safety diagnosis reports. It is designed to operate with enable_thinking=True and reasoning_effort="medium" for enhanced reasoning capabilities.

Key Capabilities

  • Improved Facility Assessment Accuracy: Achieves 70% accuracy in 6-level facility judgment, a 10% improvement over its predecessor (v1).
  • Robust Output Formatting: Consistently produces outputs with normal thinking termination (</think>) and adheres strictly to JSON schemas, achieving 100% success in both metrics during evaluation.
  • General Reasoning Preservation: Maintains strong general reasoning performance, as indicated by an 81.7% accuracy on the AIME 2024+2025 benchmark, suggesting that domain-specific fine-tuning did not significantly degrade broader capabilities.
  • Structured JSON Generation: Capable of generating structured JSON outputs that include supporting clauses and report citations.

Intended Use Cases

  • Auxiliary Review: Supports the review of facility reports for compliance with guidelines.
  • Drafting and Quality Control: Assists in drafting review reports and quality control processes.

Limitations

  • Performance is evaluated on a synthetic dataset of 50 cases and may not generalize to full, real-world reports.
  • The model's performance on strict judgment categories with limited examples requires further evaluation.
  • It may generate facts or grounds not present in the original report, necessitating cross-verification with source documents.

Finetunes
1 models