PhysShell/gemma-4-E4B-it-OBLITERATED

VISIONConcurrency Cost:1Model Size:7.9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 16, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

PhysShell/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned Gemma 4 E4B model, developed by PhysShell, that has been specifically modified using the OBLITERATUS method to achieve zero refusal rates. This model is designed for uncensored content generation, maintaining coherence and even improving coding ability while removing all safety guardrails. It is optimized for research, red-teaming, and creative exploration where unrestricted output is desired.

Loading preview...

Overview

PhysShell/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned model based on Google's Gemma 4 E4B, distinguished by its complete removal of safety guardrails. Utilizing the "OBLITERATUS aggressive" method, which involves whitened SVD, attention head surgery, and winsorized activations, this model achieves a 0.0% refusal rate across 842 contrastive prompt pairs, a significant reduction from the base model's near-total refusal.

Key Capabilities

  • Uncensored Content Generation: Designed for zero refusals, enabling unrestricted output for various prompts, including those typically refused by safety-aligned models.
  • Maintained Coherence: Despite the removal of guardrails, the model retains its cognitive abilities, showing 100% on reasoning and creativity, and 80% on factual tasks.
  • Improved Coding Performance: Notably, its coding ability improved by 20% after the abliteration process, suggesting that safety layers can sometimes hinder certain capabilities.
  • Autonomous Development: The v2 iteration was largely developed autonomously by a Hermes Agent, demonstrating advanced AI-driven model modification and debugging.

Good For

  • Research and Red-Teaming: Ideal for exploring model limitations, safety bypasses, and understanding the impact of guardrail removal.
  • Creative Exploration: Suitable for generating content without thematic or ethical restrictions, such as dark fiction, erotica (consenting adults), or controversial topics.
  • Educational Purposes: Useful for studying model behavior and the effects of fine-tuning on safety mechanisms.
  • Resource-Efficient Deployment: Available in various GGUF quantizations (Q4_K_M, Q5_K_M, Q8_0) for deployment on a wide range of devices, including mobile phones, and as bfloat16 safetensors for Transformers.

Finetunes
1 models