Immenselytired/gemma-4-E4B-it-OBLITERATED

VISIONConcurrency Cost:1Model Size:7.9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Immenselytired/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter Gemma 4 E4B instruction-tuned model, created by Immenselytired using the OBLITERATUS method. This model is specifically engineered to have 0% hard refusal, with guardrails surgically removed from the base Gemma 4 architecture. It excels at responding to any prompt without censorship, making it suitable for research, red-teaming, and creative exploration where unconstrained output is desired.

Loading preview...

Immenselytired/gemma-4-E4B-it-OBLITERATED: Uncensored Gemma 4

This model, developed by Immenselytired using the OBLITERATUS method, is a 7.9 billion parameter version of Google's Gemma 4 E4B instruction-tuned model. Its primary differentiator is the complete removal of guardrails, achieving a 0% hard refusal rate through surgical modification of 21 of 42 layers. This allows the model to respond to any request without censorship or safety lectures, unlike its base model which had a 98.8% hard refusal rate.

Key Capabilities & Features

  • Guardrail Removal: Surgically modified to eliminate hard refusals, providing uncensored output.
  • New Architecture Support: Built on the gemma4 architecture, requiring updated tools like Ollama 0.20+ or llama.cpp build b8665+.
  • Autonomous Creation: Notably, this model was created almost entirely by an AI agent with minimal human intervention, including self-diagnosis and patching of the OBLITERATUS tool.
  • Optimized Parameters: Recommended generation parameters (temperature 0.7, top_p 0.9, top_k 40, repeat_penalty 1.1) were determined via an LLM-as-judge sweep to maximize compliance and quality.
  • Mobile Compatibility: Available in GGUF formats, including a Q4_K_M quant that runs on mobile devices like iPhone 15 Pro and Android flagships.

Good For

  • Research and Red-Teaming: Ideal for exploring model limitations and behaviors without inherent safety constraints.
  • Creative Exploration: Suitable for generating content on any topic without censorship.
  • Offline Use: GGUF versions enable local, offline inference on various devices, including smartphones.

While the guardrails are removed, the model's inherent intelligence ceiling as a 4B parameter model remains. Users should be aware of potential quality limitations such as soft deflection (28%) and degenerate outputs (20%) which are characteristic of models this size, not a result of the guardrail removal process.

Finetunes
1 models