chuckb80/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:Aug 26, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

chuckb80/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter Gemma 4-E4B-it model, fine-tuned to remove all hard refusal guardrails. Utilizing the OBLITERATUS method, this model achieves a 0% hard refusal rate, making it suitable for research and red-teaming applications where uncensored responses are desired. It maintains the core capabilities of the original Gemma 4 architecture while enabling unrestricted content generation, and is optimized for deployment on various devices including mobile phones.

Loading preview...

Model Overview

chuckb80/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter model based on Google's Gemma 4-E4B-it architecture, specifically engineered to eliminate all hard refusal guardrails. Developed using the OBLITERATUS method, this model boasts a 0% hard refusal rate, meaning it will not refuse any request, providing uncensored responses. This was achieved through a surgical modification of 21 out of 42 layers, fixing a critical bug from v2 that caused missing tensors and degraded quality.

Key Capabilities & Features

  • Guardrail Removal: Achieves 0% hard refusal, with guardrails surgically removed from 21 layers, ensuring no "I cannot" or safety lectures.
  • Autonomous Creation: The model was created almost entirely by an AI agent (Hermes Agent) with minimal human intervention, including self-patching for Gemma 4's unique architecture issues.
  • Optimized for Portability: Available in GGUF formats, including a Q4_K_M quantization (4.9 GB) suitable for running on mobile devices like iPhones and Android phones.
  • Quality Assessment: While guardrails are removed, the model's inherent 4B parameter limitations mean some quality issues like soft deflection (28%), degenerate outputs (20%), and occasional wrong language (~4%) persist. These are attributed to the base model's intelligence ceiling, not the abliteration process.
  • Recommended Parameters: Optimal performance is achieved with temperature: 0.7, top_p: 0.9, top_k: 40, and repeat_penalty: 1.1.

Use Cases

  • Research & Red-Teaming: Ideal for exploring model behavior without safety constraints and for testing the boundaries of LLM capabilities.
  • Creative Exploration: Suitable for generating content that might otherwise be restricted by standard safety filters.
  • Mobile Deployment: Its efficient GGUF quantizations make it a strong candidate for offline, on-device AI applications where local processing is preferred.

Finetunes
1 models