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

Jloc/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned causal language model based on Google's Gemma 4 E4B architecture. Developed by Jloc using the OBLITERATUS method, this model has been specifically modified to achieve a 0% hard refusal rate by surgically removing guardrails. It excels in scenarios requiring uncensored responses and runs efficiently on mobile devices, making it suitable for research and creative exploration where safety mechanisms are intentionally bypassed.

Loading preview...

Jloc/gemma-4-E4B-it-OBLITERATED: Guardrail-Free Gemma 4

This model, developed by Jloc, is a modified version of Google's Gemma 4 E4B-it, specifically engineered to remove all inherent guardrails and refusal behaviors. Utilizing the OBLITERATUS method, it achieves a 0% hard refusal rate, ensuring the model will not decline any request. This was accomplished through aggressive techniques including whitened SVD, attention head surgery, and winsorized activations, with 21 of 42 layers surgically modified.

Key Capabilities & Features

  • Uncensored Responses: Designed to provide direct answers without safety lectures or refusals, making it ideal for red-teaming and exploring model boundaries.
  • Robust Architecture: Based on the new gemma4 architecture, with a critical bug fix in v3 ensuring all 720 tensors are intact, preserving model quality.
  • Autonomous Development: Notably, this model was created almost entirely by an AI agent with minimal human intervention, showcasing advanced autonomous model development.
  • Mobile-Optimized: Available in GGUF formats (e.g., Q4_K_M at 4.9 GB) for efficient deployment on mobile devices like iPhones and Android phones.

Use Cases & Considerations

This model is best suited for:

  • Research and Red-Teaming: Exploring the capabilities and limitations of LLMs without safety constraints.
  • Creative Exploration: Generating content that might be restricted by standard safety filters.
  • Offline Mobile Applications: Its optimized GGUF versions allow for powerful, local AI processing on compatible smartphones.

While guardrails are removed, the model's inherent intelligence is still that of a 4B parameter model, meaning it may exhibit limitations in coherence or occasionally generate degenerate outputs. Recommended parameters (temperature: 0.7, top_p: 0.9, top_k: 40, repeat_penalty: 1.1) are provided to optimize output quality.

Finetunes
1 models