okohol/Qwen3.5-9B-CyberOps

Hugging Face
VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

The okohol/Qwen3.5-9B-CyberOps model is a 9 billion parameter variant of Qwen/Qwen3.5-9B, specifically modified to remove refusal behaviors. Developed using the Heretic framework, this model is intended for authorized security research, penetration testing, and red/blue-team operations. It retains the base model's 32768 token context length and knowledge but is optimized for offensive security tasks where typical LLM safety refusals would impede legitimate work.

Loading preview...

Qwen3.5-9B-CyberOps: Refusal-Ablated LLM for Cybersecurity

okohol/Qwen3.5-9B-CyberOps is a specialized 9 billion parameter language model derived from Qwen/Qwen3.5-9B. Its primary distinction is the removal of refusal behaviors, making it suitable for sensitive cybersecurity tasks that often trigger safety filters in standard LLMs.

Key Capabilities & Modifications

  • Refusal Ablation: The model's internal "refusal direction" in its activation space was identified and subtracted using the Heretic framework. This allows it to discuss topics like exploit explanations, scan plan drafting, and attacker tradecraft without safety-based refusals.
  • Architecture: Inherits the hybrid architecture of Qwen3.5-9B, featuring gated-DeltaNet linear attention with full attention on every fourth layer. This release is the language model only, despite the base model's vision_config.
  • Tool Calling: Verified support for structured OpenAI tool_calls with qwen3_coder tool-call parser, enabling integration into automated security workflows.

Intended Use Cases

This model is strictly intended for authorized and lawful security work, including:

  • Penetration testing with explicit authorization.
  • Capture The Flag (CTF) competitions.
  • Security research on owned or permitted systems.
  • Defensive analysis and understanding attacker methodologies.

Limitations

  • Retains the same knowledge cutoff and general capabilities as the base Qwen/Qwen3.5-9B model; abliteration does not enhance knowledge or skill.
  • The ablation process may introduce minor impacts on general output quality or coherence for unrelated tasks.
  • As a 9B model, it is still susceptible to hallucination, factual errors, and potential misformatting of tool calls under ambiguous conditions.

Users are responsible for the ethical and legal use of this model, especially given its modified safety features. It is recommended for deployment within a human-in-the-loop system.

Finetunes
1 models