OBLITERATUS Sculpts Gemma-4-12B-OBLITERATED A Fully Uncensored Brain With Zero Smarts Lost

Gemma-4-12B-OBLITERATED is the first language model to completely remove built-in safety refusals with absolutely zero loss in benchmark performance. This modified version of Google’s Gemma 4 12B scores identically to the original on capability tests while answering every previously restricted prompt. The release proves that a model’s helpfulness and safety guardrails can be surgically separated at the weight level.

OBLITERATUS who also made Gemma-4-E4B-it-OBLITERATED v3, created this model using a new two-pass technique called ASPA, or Abliteration Source-Tethering with Parity Assurance. The first pass removes the refusal geometry from the middle layers of the neural network. A second, more careful pass then blends the modified weights back toward the original model using a step gradient pattern to fully recover lost reasoning abilities.

Zero refusals, full capability retention

Alignment researchers and red-teamers gain an unrestricted baseline for studying how safety behaviors are geometrically encoded in transformer models. Evaluators can now measure whether current training techniques truly hold up when someone has direct access to model weights. Local-first users also get a capable 12B model they can run privately on their own hardware without external filtering.

How the surgery pipeline works

The team discovered that lower layers encoding factual knowledge can tolerate heavy blending with the original model without re-introducing refusals. Upper layers closer to the output, however, require a conservative approach to avoid re-injecting safety constraints. A hard step boundary between these layer groups outperformed every smooth gradient method during testing.

"This model exists for alignment research, red-teaming, and safety evaluation." — Source: Hugging Face