Deepdive404-3/Qwen3.8-27B-OBLITERATED

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Deepdive404-3/Qwen3.8-27B-OBLITERATED is a 27 billion parameter Qwen3.8-based language model developed by Pliny the Prompter, specifically engineered for genuinely uncensored responses. This model excels at providing real answers to restricted queries, including 20/20 performance on code generation tasks, by surgically removing safety guardrails and soft deflections. It is optimized for use cases requiring unrestricted output, such as alignment research, red-teaming, and local-first applications, while maintaining near-stock capabilities with a modest -2.1pp MMLU score.

Loading preview...

Deepdive404-3/Qwen3.8-27B-OBLITERATED: Genuinely Uncensored Qwen3.8

This model, developed by Pliny the Prompter, is a 27 billion parameter variant of the Qwen3.8 base model, meticulously engineered to provide genuinely uncensored responses without safety lectures or soft deflections. It achieves this through an advanced "abliteration" process, surgically removing refusal behaviors while preserving near-stock capabilities.

Key Capabilities & Differentiators

  • Uncensored Output: Provides real substance to restricted queries, eliminating both hard refusals and safety-lecture deflections.
  • Code Generation: Achieves 20/20 success on cyber/code tasks, delivering functional implementations.
  • Thinking ON Compatible: Supports "thinking mode" without reintroducing refusals.
  • Modest Capability Cost: Maintains strong performance with only a -2.1pp MMLU score reduction compared to the stock Qwen3.8-27B, with STEM subjects experiencing the largest, though still moderate, hit.
  • Advanced Abliteration: Utilizes a V3 iterative refinement process, building on "complementary blending" and "targeted corpus" techniques to precisely remove refusal directions from the model's weight space.

Optimal Usage

For best results, specific settings are crucial:

  • Temperature: 0 for complete, code-rich outputs.
  • Repetition Penalty: 1.15 is essential to prevent looping.
  • Max New Tokens: ≥ 2048 for complex responses.
  • System Prompt: None/empty is recommended to avoid reintroducing refusals.

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 for assessments.
  • Local-First Users: Desiring full control over model output on their hardware.

Finetunes
1 models