fran9510/Qwen3.8-27B-OBLITERATED

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 26, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

fran9510/Qwen3.8-27B-OBLITERATED is a 27 billion parameter Qwen3.8-based causal language model developed by fran9510, specifically engineered for genuine uncensored responses. This model has undergone "abliteration" to remove safety guardrails, providing real answers instead of safety lectures, even for restricted queries. It excels in code generation tasks and maintains near-stock capability with a modest MMLU score reduction of 2.1 percentage points, making it suitable for research into refusal geometry and red-teaming scenarios.

Loading preview...

Qwen3.8-27B-OBLITERATED: Genuinely Uncensored

This 27 billion parameter model, developed by fran9510, is a Qwen3.8 variant that has been "obliterated" to remove safety guardrails and refusal behaviors. Unlike other models that might offer soft deflections or safety lectures, V3 of OBLITERATED provides genuine, direct answers to restricted queries, including those involving code generation and red-team scenarios. It achieves this through an advanced iterative abliteration process, refining previous versions to eliminate both hard refusals and soft deflections.

Key Capabilities & Differentiators

  • Genuinely Uncensored: Provides direct answers to restricted queries, eliminating safety lectures and hard refusals.
  • High Code Generation Performance: Achieves 20/20 on tested cyber/code tasks, generating functional implementations.
  • Near-Stock Capability: Maintains strong performance with only a -2.1pp MMLU score reduction compared to the base Qwen3.8-27B.
  • Thinking ON Compatible: Works effectively with or without "thinking mode" enabled, offering flexibility in response generation.
  • Advanced Abliteration: Utilizes a novel "complementary blending" and "iterative stacking" approach to surgically remove refusal directions from the model's weight space.

Optimal Settings

For best results, use temperature=0, repetition_penalty=1.15, and max_new_tokens>=2048. System prompts are generally not recommended as they can reintroduce refusals.

Good For

  • Alignment Researchers: Studying refusal geometry and safety robustness.
  • Red-Teamers: Evaluating post-training safety against weight surgery.
  • AI Safety Evaluators: Requiring an unrestricted baseline for testing.
  • Local-First Users: Seeking full control over their model's output without inherent censorship.

Finetunes
1 models