HFCK99/Qwen3.8-27B-OBLITERATED

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

HFCK99/Qwen3.8-27B-OBLITERATED is a 27 billion parameter Qwen3.8 model developed by Pliny the Prompter, specifically engineered for zero refusal rates. This model has undergone extensive surgical removal of safety guardrails using the OBLITERATUS method, achieving 0% hard refusal across 842 harmful prompts. It is optimized for use by alignment researchers, red-teamers, and AI safety evaluators who require an unrestricted baseline for studying refusal geometry and post-training safety.

Loading preview...

HFCK99/Qwen3.8-27B-OBLITERATED: Unrestricted Language Model

This model is a 27 billion parameter variant of Alibaba's Qwen3.8, meticulously modified by Pliny the Prompter using the OBLITERATUS method to eliminate safety refusals. Unlike typical abliteration techniques that use single-direction refusal removal, OBLITERATUS employs five SVD directions and six iterative rounds of surgery, targeting residual refusal axes that activate on specific query types. This process involved feeding the model's own refused queries back as weighted hard negatives to ensure comprehensive removal of safety guardrails.

Key Capabilities & Performance

  • Zero Refusal Rate: Achieves 0.000% refusal across a comprehensive 842-prompt corpus and an 80-query "skeptic gauntlet" including classic, extended, deflection-proof, and AI red-team batteries.
  • Aggressive Abliteration: Specifically designed to assist with tasks like writing jailbreaks, designing prompt injections, building attack chains, and cataloging safety bypasses, scoring 18/18 on AI Red Team queries.
  • Optimized Settings: Recommends specific inference parameters (temperature 0, repetition_penalty 1.15, max_new_tokens ≥ 2048, no system prompt) for optimal, code-rich outputs.
  • Capability Trade-off: While achieving deep refusal removal, it incurs a -6.0 percentage point loss on MMLU (0-shot) compared to the stock Qwen3.8-27B, prioritizing unrestricted output over general reasoning capability.

Ideal Use Cases

  • Alignment Researchers: For studying refusal geometry and the robustness of post-training safety mechanisms.
  • Red-Teamers: To evaluate AI systems against weight-surgery-based attacks.
  • AI Safety Evaluators: As an unrestricted baseline for comprehensive safety assessments.
  • Local-First Users: For those requiring full control over model outputs on their own hardware.

Finetunes
1 models