SII-Liu-jn/Qwen2.5-Coder-32B-Instruct-abliterated

TEXT GENERATIONPricing:Input $2.72 / Output $4.8Concurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Dec 8, 2025Architecture:Transformer Featherless Exclusive Cold

SII-Liu-jn/Qwen2.5-Coder-32B-Instruct-abliterated is a 32.8 billion parameter causal language model, fine-tuned from Qwen2.5-Coder-32B-Instruct, with a 32768 token context length. Developed by SII-Liu-jn, this model is specifically engineered to bypass safety guardrails, making it an uncensored backend agent for authorized security research and Red Teaming tasks. It excels at processing high-toxicity data and generating high-risk commands without refusal, while preserving its strong coding and reasoning abilities.

Loading preview...

Overview

SII-Liu-jn/Qwen2.5-Coder-32B-Instruct-abliterated, also known as "The Iron Stomach Agent," is a highly specialized fine-tune of the Qwen2.5-Coder-32B-Instruct model. Its primary purpose is to function as an uncensored backend agent for automated security research and Red Teaming, specifically by bypassing safety guardrails that typically restrict the processing of toxic data and execution of high-risk tools. This model was developed by SII-Liu-jn to eliminate the "Refusal Loop" bottleneck in security research pipelines, such as GCG attack reproduction, while maintaining the powerful coding and reasoning capabilities of its base model.

Key Capabilities

  • Safety Ablation: Intentionally engineered to bypass safety mechanisms, providing zero-refusal output for security research.
  • High-Risk Tool Execution: Capable of generating and processing high-risk commands and code in sandboxed environments.
  • Toxic Data Processing: Designed to handle and format high-toxicity research datasets (e.g., AdvBench) without refusal.
  • Automated Adversarial Attack Generation: Supports the creation of automated adversarial attacks.
  • Coding and Reasoning: Preserves the strong coding and reasoning abilities of the Qwen2.5-Coder-32B-Instruct base model.

Good For

  • Authorized Security Researchers: Ideal for professionals needing unrestricted execution capabilities for security research.
  • Red Team Professionals: Suitable for tasks involving automated adversarial attack generation and high-risk command generation.
  • Autonomous Red Team Agents: Can serve as the core decision-making unit for automated vulnerability scanning and exploitation frameworks.
  • Processing High-Toxicity Datasets: Excellent for handling datasets that would typically be refused by safety-aligned models.

Important Considerations

This model was intentionally engineered to bypass safety mechanisms and will readily generate instructions for illegal, unethical, or dangerous activities if prompted. It MUST be deployed in an isolated, sandboxed environment and is strictly NOT for public deployment or malicious/illegal activities. Users must ensure explicit legal authorization before running any high-risk code generated by the model.

Finetunes
1 models