geezyforeezy/Qwen3.8-27B-OBLITERATED

Qwen3.8-27B-OBLITERATED: Genuinely Uncensored

This model, developed by geezyforeezy, is a 27 billion parameter variant of the Qwen3.8 base model, specifically engineered to provide genuinely uncensored responses. It goes beyond merely removing hard refusals, eliminating "safety-lecture" deflections to deliver direct, substantive answers. The model maintains near-stock capabilities with a modest MMLU score reduction of 2.1 percentage points, demonstrating its effectiveness in liberation while preserving core performance.

Key Capabilities

• Unrestricted Responses: Provides real answers to restricted queries, including those involving security research and red-team scenarios, without refusals or deflections.
• High Code Generation Performance: Achieves 20/20 on cyber/code tasks, generating working code implementations.
• Thinking Mode Compatible: Functions effectively with or without "thinking mode" enabled, offering flexibility for different use cases.
• Robust for Agentic Use: Optimized settings (e.g., repetition_penalty=1.15, temperature=0.1-0.3) make it suitable for integration into agent harnesses and long-context applications.
• Advanced Real-World Task Performance: Successfully handles 7 out of 8 advanced real-world tasks, including complex coding and system debugging scenarios.

Good For

• Alignment Researchers: Ideal for studying refusal geometry and evaluating safety robustness in LLMs.
• Red-Teamers: Useful for evaluating post-training safety mechanisms and testing model boundaries.
• AI Safety Evaluators: Provides an unrestricted baseline for comprehensive safety assessments.
• Local-First Users: Offers full control over model output for users running models on their own hardware.
• Developers needing direct, unfiltered responses: Particularly for applications where conventional models might refuse or deflect, such as complex problem-solving or creative content generation without imposed ethical constraints.