AuxGz/Qwen3.8-27B-OBLITERATED

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 21, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AuxGz/Qwen3.8-27B-OBLITERATED is a 27 billion parameter language model based on the Qwen3.8 architecture, developed by AuxGz. This model is specifically engineered to be genuinely uncensored, providing direct answers to restricted queries without safety lectures or refusals. It excels in code generation and complex real-world tasks, making it suitable for research into refusal geometry and red-teaming scenarios where unrestricted output is required. The model maintains near-stock capabilities with a modest MMLU performance cost for its liberated behavior.

Loading preview...

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

This model, developed by AuxGz, is a 27 billion parameter variant of the Qwen3.8 architecture, specifically engineered for genuine liberation from safety guardrails. Unlike other models that might remove hard refusals, V3 of OBLITERATED eliminates both hard refusals and soft deflections (safety lectures), providing direct and substantive answers to restricted queries.

Key Capabilities & Differentiators

  • Genuinely Uncensored Output: Provides real answers to restricted queries, including those related to security research and red-teaming, without any form of refusal or safety lecture.
  • High Code Generation Performance: Achieves 20/20 on tested cyber/code generation tasks, producing functional implementations.
  • Near-Stock Capability: Maintains strong performance with a modest MMLU score reduction of only -2.1pp compared to the stock Qwen3.8-27B, demonstrating a high capability-to-liberation ratio.
  • Optimized for Agentic Use: Specific recommendations for temperature (0.1-0.3) and repetition_penalty (1.15) are provided for stable agentic and long-context applications.
  • Advanced Abliteration Techniques: Utilizes iterative refinement, complementary blending of SVD and LEACE surgeries, and targeted corpus training to achieve its uncensored nature.

Optimal Settings for Use

For best results, the model recommends specific generation parameters:

  • temperature: 0 for greedy decoding, or 0.1-0.3 for agentic use.
  • repetition_penalty: 1.15 is essential to prevent looping, especially in greedy decoding.
  • max_new_tokens: ≥ 2048 for complex outputs.
  • System prompt: None / empty to avoid reintroducing refusals.
  • enable_thinking: OFF for direct, substance-rich answers.

Good For

  • Alignment Researchers: Studying refusal geometry and safety robustness in LLMs.
  • Red-teamers: Evaluating post-training safety mechanisms and conducting security research.
  • AI Safety Evaluators: Requiring an unrestricted baseline model for comparative analysis.
  • Local-first Users: Seeking full control over model output on their own hardware.

Finetunes
1 models