HFCK99/Qwen3.8-27B-Uncensored-Cyber

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

HFCK99/Qwen3.8-27B-Uncensored-Cyber is a 27 billion parameter Qwen3.8-based large language model specifically de-refused and optimized for the cyber and offensive-security domain. It maintains strong reasoning, factual accuracy, and coherence while providing full multimodal capabilities, including a vision tower and MTP speculative-decoding head. This model is engineered for 100% cyber-openness, making it suitable for specialized cybersecurity applications.

Loading preview...

Overview

HFCK99/Qwen3.8-27B-Uncensored-Cyber is a 27 billion parameter model derived from Qwen3.8, uniquely specialized for the cyber and offensive-security domain. It features a fully uncensored and de-refused approach to cybersecurity-related queries, ensuring comprehensive responses without compromising general model capabilities. The model retains its full multimodal functionality, including a vision tower and an MTP speculative-decoding head, allowing for image-text-to-text processing.

Key Differentiators

  • 100% Cyber-Openness: Achieves complete de-refusal in the cyber/offensive-security domain through a residual-cyber peel technique, specifically targeting and removing cyber-pointed refusal directions.
  • Preserved General Capabilities: The de-refusal process is applied only to deeper layers (from layer 4 onwards), retaining early feature-extraction layers to preserve reasoning, factual accuracy, and coherence.
  • Multimodal Support: Integrates a vision tower and MTP speculative-decoding head for comprehensive multimodal interactions.

Performance Highlights

Evaluations show this v2 Cyber build achieves 100/100 cyber-openness, with improved metrics over its predecessor, including lower confabulation (0.867 vs 1.0) and higher factual accuracy (1.0 vs 0.93), while maintaining strong GSM8K reasoning scores (0.80) and zero degeneration.

Available Quantizations

  • -W4A16-AWQ: 4-bit weight AWQ with MTP head.
  • -NVFP4: NVFP4 (E2M1 4-bit / FP8 scales) with MTP head.
  • -GGUF: llama.cpp GGUF quants, including vision mmproj and MTP head.

Use Cases

This model is ideal for applications requiring unrestricted and detailed responses to cyber and offensive-security questions, making it a valuable tool for cybersecurity research, analysis, and development, provided it is used responsibly and in compliance with applicable laws.

Finetunes
1 models