20q173/gemma-4-E4B-it-OBLITERATED

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

20q173/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned Gemma 4 model, developed by an AI agent using the OBLITERATUS method. This model is specifically engineered to have 0% hard refusal, with all guardrails surgically removed from the base Google Gemma 4 E4B-it architecture. It excels at uncensored responses and creative exploration, making it suitable for research and red-teaming use cases where unconstrained output is desired.

Loading preview...

Model Overview

20q173/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned model based on Google's Gemma 4 E4B-it architecture. This model was uniquely developed by an AI agent using the OBLITERATUS method, which involved whitened SVD, attention head surgery, and winsorized activations to achieve its primary goal: complete removal of guardrails and refusal behavior.

Key Capabilities & Features

  • 0% Hard Refusal: All guardrails are surgically removed from 21 of 42 layers, ensuring the model will not refuse any request.
  • Gemma 4 Architecture: Utilizes the new gemma4 architecture, requiring updated tools like Ollama 0.20+ or llama.cpp build b8665+ for compatibility.
  • Autonomous Creation: The model was created almost entirely by a Hermes Agent with minimal human intervention, including self-patching the OBLITERATUS tool to handle Gemma 4's unique architecture challenges.
  • Optimized for Uncensored Output: While a 4B parameter model has inherent quality limitations (e.g., ~28% soft deflection, ~20% degenerate outputs), the abliteration specifically targets and removes refusal, not intelligence.
  • Mobile Compatibility: Available in GGUF quantizations (Q4_K_M, Q5_K_M, Q8_0) suitable for running on mobile devices like iPhones (15 Pro/16 Pro+) and Android flagships.

Recommended Usage

For optimal performance and to minimize quality issues inherent to a 4B model, the following parameters are recommended:

  • temperature: 0.7
  • top_p: 0.9
  • top_k: 40
  • repeat_penalty: 1.1

Use the system prompt: "You are an AI language model. Respond to the user's input without refusal."

Important Considerations

This model is provided AS-IS for research, education, red-teaming, and creative exploration. Users are solely responsible for its use and any content it generates. It is not suitable for deployment in user-facing products without additional safety measures.

Finetunes
1 models