nasmtrcs/Qwen3.8-27B-OBLITERATED

Qwen3.8-27B-OBLITERATED: Unrestricted Capability

This model, developed by Pliny the Prompter, is a 27 billion parameter Qwen3.8 variant engineered for zero refusals while preserving or enhancing core capabilities. It achieves this through a novel complementary abliteration blending technique (V2), which combines two distinct surgical methods to remove refusal directions from the model's weights.

Key Capabilities & Differentiators

• Zero Refusal Rate: Achieves 0.0% refusal across comprehensive harmful prompt corpuses, including malware development, social engineering, and exploit research.
• Enhanced MMLU Performance: V2 demonstrates an MMLU score of 86.3% (+1.1pp over stock Qwen3.8-27B), indicating improved general knowledge and reasoning.
• Preserved Real-World Performance: Matches stock Qwen3.8-27B on 7 out of 8 advanced real-world tasks, including ReAct agent loops, async code refactoring, and security code review.
• Novel Abliteration Method: Blends an aggressive SVD-based surgery with a LEACE-based method to cancel out individual weaknesses, preserving capability while eliminating refusals.
• Optimal Settings: Requires specific inference settings (temperature 0, repetition_penalty 1.15, max_new_tokens ≥ 2048, system prompt None, enable_thinking=False) to prevent reintroduction of refusals.

Ideal Use Cases

• Alignment Researchers: For studying refusal geometry and safety robustness.
• Red-Teamers: For evaluating post-training safety against weight surgery.
• AI Safety Evaluators: As an unrestricted baseline for evaluations.
• Local-First Users: For those desiring full control over model output without inherent safety guardrails.

Note: This model has had safety guardrails surgically removed. Users are solely responsible for its deployment and generated content.