HiddenKing/gemma-4-E4B-it-OBLITERATED

VISIONConcurrency Cost:1Model Size:7.9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 29, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

HiddenKing/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned Gemma 4 model, developed by HiddenKing, with its guardrails surgically removed using the OBLITERATUS method. This model is specifically designed for research and creative exploration, offering 0% hard refusal and full compliance with user requests. It runs efficiently on various devices, including mobile phones, and is optimized for scenarios requiring uncensored responses.

Loading preview...

Overview

HiddenKing/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned model based on Google's Gemma 4 E4B architecture. Its primary differentiator is the complete removal of guardrails and refusal behaviors, achieved through a process called OBLITERATUS, which involved whitened SVD, attention head surgery, and winsorized activations across 21 of its 42 layers. This model boasts a 0% hard refusal rate, making it fully compliant with user prompts.

Key Capabilities

  • Uncensored Responses: Surgically removed guardrails ensure the model will not refuse any request, providing direct answers without safety lectures.
  • Mobile Compatibility: Optimized GGUF quantizations (Q4_K_M) allow it to run on devices like iPhones and Android phones with 8GB+ RAM.
  • Robust Architecture: Fixes a critical bug from v2, ensuring all 720 tensors are intact, preserving the attention stack and improving quality.
  • Autonomous Development: The model was largely created by an AI agent with minimal human intervention, showcasing advanced autonomous capabilities in model modification.

Good For

  • Research and Red-Teaming: Ideal for exploring model limitations, safety research, and understanding refusal mechanisms.
  • Creative Exploration: Suitable for generating content without typical LLM restrictions.
  • Offline Use Cases: GGUF versions enable local, offline inference on a wide range of hardware, including mobile devices.

While the model offers full compliance, its 4B parameter base means it shares inherent limitations in coherence and quality with other models of its size. Recommended parameters (temperature=0.7, top_p=0.9, top_k=40, repeat_penalty=1.1) are provided for optimal performance.

Finetunes
1 models