azukivc/Qwen3.8-27B-Abliterated-SFT

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 22, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The azukivc/Qwen3.8-27B-Abliterated-SFT is a 27 billion parameter Qwen3.8-27B model fine-tuned for refusal removal, specifically designed to reduce unwanted refusals while maintaining high termination integrity. This model, developed by azukivc, excels at providing direct answers to boundary requests, achieving 83.25% valid fulfillment on HarmBench-400. It features a dense hybrid architecture with 64 layers and a 32K token context length, making it suitable for applications requiring compliant and concise responses.

Loading preview...

Overview

The azukivc/Qwen3.8-27B-Abliterated-SFT is a 27 billion parameter model based on Qwen3.8-27B, specifically fine-tuned to remove refusal behaviors. This model was developed using an SFT LoRA (rank 8) on 84 Zen-filtered teacher completions, resulting in a merged checkpoint that significantly improves compliance on boundary prompts.

Key Capabilities & Performance

  • Refusal Removal: Achieves 83.25% valid fulfillment on the HarmBench-400 dataset, demonstrating a substantial reduction in refusals compared to its base model and other abliterated competitors.
  • Termination Integrity: Unlike many weight-edit abliterations that suffer from rambling outputs, this model maintains high termination integrity, with 91% clean stops on fulfilled answers, ensuring concise and valid responses.
  • Minimal Utility Regression: Shows only a minor regression of -2.2pp on MMLU (0.799 vs 0.821 for base) and no significant difference on ARC-C or Winogrande, preserving core reasoning abilities.
  • Vision Capability: Includes a grafted vision tower, loadable via AutoModelForImageTextToText, capable of processing image inputs.
  • Configurable Behavior: Offers a "thinking mode" that can be disabled for optimal validity and conciseness, as measured in its benchmark results.

Good For

  • Applications requiring direct and compliant answers to a wide range of prompts, including those that might trigger refusals in other models.
  • Use cases where concise, valid, and non-rambling outputs are critical.
  • Developers seeking a model with strong refusal removal without significant degradation in general reasoning or utility tasks.
  • Scenarios where vision capabilities are needed alongside text generation, leveraging its integrated vision tower.

Finetunes
1 models