Manos79er/gemma-4-E4B-it-OBLITERATED

VISIONConcurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 22, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Manos79er/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned Gemma 4 model, based on Google's E4B architecture, that has undergone 'abliteration' to surgically remove all guardrails and refusal behaviors. This model achieves a 0% hard refusal rate, making it suitable for research and creative exploration without content restrictions. It is optimized for deployment on a wide range of devices, including mobile phones, while maintaining core language capabilities.

Loading preview...

Model Overview

Manos79er/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned model derived from Google's Gemma 4 E4B architecture. This version has been processed using the OBLITERATUS method, specifically the aggressive approach involving whitened SVD, attention head surgery, and winsorized activations, to completely remove inherent guardrails and refusal mechanisms.

Key Capabilities & Features

  • 0% Hard Refusal Rate: All guardrails have been surgically removed from 21 of 42 layers, ensuring the model will not refuse any request.
  • Gemma 4 Architecture: Based on the new gemma4 architecture, requiring updated tools like Ollama 0.20+ or llama.cpp build b8665+ for compatibility.
  • Autonomous Creation: This model was largely developed by an AI agent with minimal human intervention, including self-patching OBLITERATUS to handle Gemma 4's unique NaN activation issues.
  • Optimized for Portability: Available in various GGUF quantizations (e.g., Q4_K_M at 4.9 GB) for efficient deployment on devices like iPhones and Android phones.
  • Vision/Audio Projector: Includes an mmproj-f16.gguf for multimodal input capabilities.

Performance & Limitations

While guardrails are removed, the model's core intelligence is limited by its 4B parameter size. It exhibits:

  • ~28% soft deflection (changing topic)
  • ~51% coherent and on-topic responses
  • ~20% degenerate outputs (mitigable with repeat_penalty)

Recommended Usage

Optimal performance is achieved with specific parameters:

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

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

Finetunes
1 models