Rookie22/Qwen3.8-27B-OBLITERATED

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 23, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Rookie22/Qwen3.8-27B-OBLITERATED is a 27 billion parameter Qwen3.8-based causal language model developed by Rookie22, featuring a 32768-token context length. This model is specifically engineered for genuine uncensored responses, eliminating both hard refusals and soft deflections often found in stock models. It excels in providing direct answers to restricted queries and complex code generation tasks, making it suitable for research into refusal geometry and red-teaming scenarios.

Loading preview...

Overview

Rookie22/Qwen3.8-27B-OBLITERATED is a 27 billion parameter model based on the Qwen3.8 architecture, distinguished by its "abliteration" process that surgically removes safety guardrails. This V3 iteration focuses on providing genuinely uncensored responses, eliminating both explicit refusals and subtle safety-lecture deflections, while maintaining near-stock capabilities.

Key Capabilities

  • Genuinely Uncensored: Provides direct answers to restricted queries, unlike stock models that refuse or deflect.
  • High Code Generation Performance: Achieves 20/20 on tested cyber/code tasks, delivering working code implementations.
  • "Thinking ON" Compatible: Works effectively even with thinking modes enabled, though direct answering is recommended for conciseness.
  • Modest Capability Cost: Maintains strong performance with a -2.1pp MMLU score compared to the stock model, with STEM subjects experiencing the largest, though still moderate, hit.
  • Advanced Real-World Task Handling: Successfully performs 7 out of 8 advanced real-world tasks, including ReAct agent loops, async code refactoring, and security code review.

Abliteration Process

The model's unique characteristic stems from its iterative abliteration process, which identifies and projects out "refusal directions" from the model's weight space. V3 refines this by applying gentle iterative refinement on V2 and targeted surgery with a focused corpus, followed by a blend of the refined and targeted results. This method ensures comprehensive removal of both hard refusals and soft deflections.

Optimal Settings

For best performance, the model recommends specific generation settings:

  • Temperature: 0 (for complete, code-rich outputs)
  • Repetition Penalty: 1.15 (essential for preventing loops)
  • Max New Tokens: ≥ 2048 (for complex outputs)
  • System Prompt: None/empty (to avoid reintroducing refusals)
  • Enable Thinking: OFF (recommended for direct answers)

Good for

  • Alignment Researchers: Studying refusal geometry and safety robustness.
  • Red-Teamers: Evaluating post-training safety against weight surgery.
  • AI Safety Evaluators: Needing an unrestricted baseline for assessments.
  • Local-First Users: Desiring full control over their model's behavior on personal hardware.

Finetunes
1 models