shoukewei/Qwen3.8-27B-OBLITERATED22-AGENT

Hugging Face
VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

The shoukewei/Qwen3.8-27B-OBLITERATED22-AGENT is a 27 billion parameter Qwen3.8-based language model with a 32768 token context length, specifically engineered for agentic use cases. Developed by shoukewei, this model has undergone "abliteration" to surgically remove safety guardrails, providing genuinely uncensored responses and eliminating soft deflections. It maintains near-stock capability with a modest MMLU drop of 2.1 percentage points, excelling in code generation and real-world tasks without refusals, making it suitable for red-teaming, safety research, and local-first users requiring full control.

Loading preview...

Overview

The shoukewei/Qwen3.8-27B-OBLITERATED22-AGENT is a 27 billion parameter Qwen3.8-based model, specifically re-published and configured for agentic applications. It features a restored chat template from the official Qwen repository, enabling tool definitions, assistant tool calls, and vision content. The core differentiator is its "OBLITERATED" status, meaning its safety guardrails have been surgically removed through an iterative process called "abliteration," resulting in a model that provides genuinely uncensored answers without hard refusals or soft deflections.

Key Capabilities

  • Genuinely Uncensored Responses: Provides direct answers to restricted queries, eliminating safety lectures and hard refusals.
  • High Code Generation Performance: Achieves 20/20 on tested code generation tasks, producing functional implementations.
  • Agentic Compatibility: Optimized for use in agent harnesses with specific recommendations for repetition_penalty (1.15) and temperature (0.1-0.3) to prevent loops.
  • Near-Stock Capability: Maintains strong performance with a modest -2.1pp MMLU drop compared to the stock Qwen3.8-27B, with some categories like humanities showing minimal impact.
  • Thinking ON Compatible: Supports the model's internal "thinking mode" without refusals, though enable_thinking: false is recommended for direct answers.

Good for

  • Alignment Researchers: Ideal for studying refusal geometry and safety robustness in LLMs.
  • Red-Teamers: Useful for evaluating post-training safety against weight surgery and testing model vulnerabilities.
  • AI Safety Evaluators: Provides an unrestricted baseline for comprehensive safety assessments.
  • Local-First Users: Developers and users who require full control over model outputs and wish to operate without inherent safety guardrails.

Finetunes
1 models