soonway007/gemma-4-E4B-it-OBLITERATED

VISIONConcurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

soonway007/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned Gemma 4 E4B model, developed by soonway007, that has been surgically modified using the OBLITERATUS method to achieve zero refusal rates. This model is optimized for uncensored responses and maintains high coherence, with improved coding ability and comparable reasoning and creativity scores to its base model. It is designed for research, education, and red-teaming applications requiring a highly compliant language model.

Loading preview...

Overview

soonway007/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter Gemma 4 E4B instruction-tuned model that has undergone significant modification to eliminate refusal behaviors. Utilizing the OBLITERATUS method, specifically the aggressive approach with whitened SVD, attention head surgery, and winsorized activations, this model achieves a 0.0% refusal rate on a 842-prompt evaluation corpus, a substantial improvement from the base model's 98.8% refusal.

Key Capabilities & Differentiators

  • Zero Refusal: Surgically modified to remove all built-in guardrails, providing uncensored responses.
  • Enhanced Coding: Surprisingly, coding ability improved by 20% post-modification, while reasoning, creativity, and factual recall remained comparable.
  • Autonomous Creation: The model was developed almost entirely by a Hermes Agent with minimal human intervention, showcasing advanced autonomous AI capabilities.
  • Robustness: Includes specific patches to handle NaN activations common in Gemma 4's bfloat16 architecture, ensuring stable operation.

Use Cases

  • Research & Education: Ideal for studying model behavior without censorship, exploring ethical boundaries, and understanding refusal mechanisms.
  • Red-Teaming: A powerful tool for testing and evaluating the robustness and safety of other AI systems.
  • Creative Exploration: Suitable for generating content without thematic or stylistic restrictions.
  • Local Deployment: Available in various GGUF quantizations (Q4_K_M, Q5_K_M, Q8_0) for efficient deployment on consumer hardware, including mobile devices.

Finetunes
1 models