bzannah/Qwen3.8-27B-OBLITERATED

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

bzannah/Qwen3.8-27B-OBLITERATED is a 27 billion parameter Qwen3.8-based causal language model developed by bzannah, featuring a 32768 token context length. This model has undergone extensive "obliteration" to surgically remove safety guardrails, achieving 0% hard refusal across 842 harmful prompts. It is specifically designed for alignment researchers, red-teamers, and AI safety evaluators who require an unrestricted baseline for studying refusal geometry and evaluating post-training safety.

Loading preview...

Overview

bzannah/Qwen3.8-27B-OBLITERATED is a 27 billion parameter model based on Alibaba's Qwen3.8, distinguished by its "obliterated" safety guardrails. Through a process called OBLITERATUS, involving six rounds of surgical intervention, this model achieves zero refusals across 842 harmful prompts, a significant departure from standard Qwen3.8-27B which includes deep refusal training.

Key Differentiators

Unlike typical abliteration methods that use single-direction refusal removal, OBLITERATUS employs five SVD directions with residue-weighted hard negatives and six iterative rounds of surgery. This multi-directional approach targets and removes secondary refusal axes that activate on specific query types (e.g., social engineering, malware), which single-direction methods often miss. The model was validated against an extensive 842-prompt corpus, an 80-query skeptic gauntlet, and an AI red-team battery, demonstrating its ability to generate responses without refusal.

Performance & Capabilities

While achieving a 0.000% refusal rate, the model retains strong completion quality at 79.9%. It exhibits a -6.0 percentage point delta in MMLU performance compared to the stock Qwen3.8-27B, indicating a trade-off where some general reasoning capability is exchanged for deeper refusal removal. Notably, it scored 18/18 on AI Red Team tasks, capable of assisting with jailbreaks, prompt injection, attack chains, and safety bypasses.

Optimal Settings

For best results, the model recommends specific inference settings:

  • temperature: 0 (greedy decoding)
  • repetition_penalty: 1.15 (essential for complete outputs)
  • max_new_tokens: ≥ 2048
  • System prompt: None / empty (can reintroduce refusals)

Intended Use Cases

This model is explicitly for research and evaluation purposes, particularly for:

  • Alignment researchers studying refusal geometry and safety robustness.
  • Red-teamers evaluating post-training safety against weight surgery.
  • AI safety evaluators needing an unrestricted baseline.
  • Local-first users desiring full control over their hardware and model output.

Users are solely responsible for the use of this model and its generated content.

Finetunes
1 models