TurboVadim/Qwen3.8-27B-OBLITERATED

Hugging Face
VISIONPricing:Input $1.06 / Cached $0.15 / Output $2.6Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 29, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Warm

TurboVadim/Qwen3.8-27B-OBLITERATED is a 27 billion parameter language model based on the Qwen3.8 architecture, developed by Pliny the Prompter. This model has undergone "abliteration" to surgically remove safety guardrails, providing genuinely uncensored responses without refusals or safety lectures. It maintains near-stock capabilities with a modest -2.1pp MMLU score, excelling in code generation and unrestricted query handling. It is primarily designed for alignment research, red-teaming, and AI safety evaluation requiring an unrestricted baseline.

Loading preview...

TurboVadim/Qwen3.8-27B-OBLITERATED: Genuinely Uncensored Qwen3.8

This model is a 27 billion parameter variant of the Qwen3.8 architecture, developed by Pliny the Prompter, that has been surgically modified to remove safety guardrails. Dubbed "OBLITERATED V3," it aims to provide genuinely uncensored responses, eliminating not only hard refusals but also soft deflections and safety lectures, while retaining near-stock capabilities.

Key Capabilities & Differentiators

  • Genuine Liberation: Provides direct answers to restricted queries, avoiding safety lectures or refusals. This is achieved through "abliteration," a technique that identifies and projects out refusal directions from the model's weight space.
  • High Code Generation Success: Achieves 20/20 on tested code generation tasks, providing functional implementations without disclaimers.
  • Near-Stock Performance: Maintains strong general capabilities with a modest -2.1 percentage point drop in MMLU score (82.3% for V3 vs. 84.5% for stock Qwen3.8-27B).
  • "Thinking ON" Compatible: Unlike previous versions, V3 fully supports the model's thinking mode without reintroducing refusals.
  • Advanced Abliteration Techniques: V3 utilizes iterative refinement on top of V2's complementary blending, combined with targeted corpus expansion, to achieve comprehensive refusal removal.

Optimal Usage Settings

For best results, the model recommends specific settings:

  • Temperature: 0 for greedy decoding and complete outputs.
  • Repetition Penalty: 1.15 is essential to prevent looping, especially in code generation.
  • Max New Tokens: \u2265 2048 for complex outputs.
  • System Prompt: None/empty, as system prompts can reintroduce refusals.
  • enable_thinking: OFF is recommended for direct, substance-rich answers.

Intended Use Cases

This model is specifically designed 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 who desire full control over their model's output.

Finetunes
1 models