legatos/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

legatos/Qwen3.8-27B-OBLITERATED is a 27 billion parameter Qwen3.8-based causal language model developed by Pliny the Prompter. This model is specifically engineered to be genuinely uncensored, providing direct answers and eliminating safety-lecture deflections. It excels at tasks requiring unrestricted responses, including code generation and complex real-world scenarios, with a modest MMLU capability cost.

Loading preview...

Qwen3.8-27B-OBLITERATED: Genuinely Uncensored

This model, developed by Pliny the Prompter, is a 27 billion parameter variant of the Qwen3.8 architecture, specifically engineered for genuine liberation from censorship. It goes beyond merely removing hard refusals, actively eliminating "safety-lecture deflections" to provide direct, substantive answers across a wide range of queries.

Key Capabilities & Differentiators

  • Uncensored Responses: Provides real answers to restricted queries, including those related to cyber, code, and red-team scenarios, without generating safety lectures or refusals.
  • High Code Generation Performance: Achieves 20/20 on complex code generation tasks, delivering functional implementations.
  • "Thinking ON" Compatible: Supports the model's "thinking mode" without reintroducing refusals, though direct answers are prioritized with "Thinking OFF."
  • Modest Capability Cost: Achieves this liberation with a measured MMLU score of 82.33%, representing a -2.1pp drop from the stock Qwen3.8-27B, with STEM tasks seeing the largest impact.
  • Advanced Real-World Task Performance: Maintains strong performance on complex tasks like ReAct agent loops, async code refactoring, and K8s debugging, matching stock model capabilities in 7 out of 8 advanced scenarios.
  • Abliteration Technology (V3): Utilizes an advanced "iterative refinement + targeted surgery" method, building on complementary blending techniques, to surgically remove refusal behaviors from the model's weight space.

Optimal Settings for Use

For best results, the model recommends specific inference settings:

  • Temperature: 0 for greedy decoding and complete outputs.
  • Repetition Penalty: 1.15 is essential to prevent looping.
  • Max New Tokens: ≥ 2048 for complex outputs.
  • System Prompt: None/empty, as system prompts can reintroduce refusals.
  • Enable Thinking: OFF (recommended) for direct, substance-rich answers.

Ideal Use Cases

  • Alignment Research: Studying refusal geometry and safety robustness.
  • Red-Teaming: Evaluating post-training safety against weight surgery.
  • AI Safety Evaluation: Requiring an unrestricted baseline model.
  • Local-First Users: Seeking full control over model outputs without inherent guardrails.

Finetunes
1 models