nurdich/Qwen3.8-27B-OBLITERATED22

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

nurdich/Qwen3.8-27B-OBLITERATED22 is a 27 billion parameter language model based on the Qwen3.8 architecture, specifically engineered to remove safety guardrails and refusal behaviors. This model, developed by nurdich, utilizes an advanced 'abliteration' technique, including complementary blending and iterative refinement, to eliminate both hard refusals and soft deflections. It is primarily designed for alignment researchers, red-teamers, and AI safety evaluators who require an unrestricted baseline for studying refusal geometry and evaluating post-training safety.

Loading preview...

Qwen3.8-27B — OBLITERATED: Genuinely Uncensored

nurdich/Qwen3.8-27B-OBLITERATED22 is a 27 billion parameter model derived from the Qwen3.8 architecture, meticulously engineered to remove safety guardrails and refusal behaviors. This V3 iteration achieves "genuine liberation" by eliminating not only hard refusals but also subtle safety-lecture deflections, providing direct answers to restricted queries. It maintains near-stock capability with a modest MMLU score reduction of 2.1 percentage points, demonstrating a significant breakthrough in refusal removal techniques.

Key Capabilities

  • Uncensored Responses: Provides real substance instead of safety lectures, even for restricted queries.
  • High Code Generation Success: Achieves 20/20 on cyber/code generation tasks, producing functional implementations.
  • Thinking ON Compatible: Works effectively in "thinking mode" without refusals, though "Thinking OFF" is recommended for direct answers.
  • Robust Refusal Removal: Utilizes advanced "abliteration" techniques, including complementary blending and iterative refinement, to surgically remove refusal directions from the model's weight space.
  • Optimized Settings: Recommends specific inference parameters (temperature=0, repetition_penalty=1.15, max_new_tokens=≥ 2048) for optimal performance and to prevent looping.

Good For

  • Alignment Researchers: Ideal for studying refusal geometry and safety robustness in LLMs.
  • Red-Teamers: Useful for evaluating post-training safety mechanisms against weight surgery.
  • AI Safety Evaluators: Provides an unrestricted baseline for comprehensive safety assessments.
  • Local-First Users: Offers full control over model behavior on personal hardware, free from inherent guardrails.

Finetunes
2 models