Huihui abliterates Gemma 4 12B, strips safety filters from Google's multimodal model

Huihui-gemma-4-12B-it-abliterated is an uncensored variant of Google's Gemma 4 12B instruction-tuned model, with built-in safety filters and content restrictions stripped out. The 12-billion-parameter multimodal model uses BF16 tensor precision and requires approximately 16 GB of memory to run locally.

Abliteration removes safety alignment from language models by targeting specific activation patterns associated with refusal behavior, allowing the model to respond to prompts it would otherwise decline. The base Gemma 4 12B-it model was released by Google with standard safety tuning; this version preserves the underlying instruction-following capabilities while removing those guardrails. The multimodal support means it can process both text and image inputs. At 12 billion parameters in bfloat16, the weights clock in around 24 GB on disk but compress to roughly 16 GB in active memory during inference, making it feasible on consumer GPUs with 24 GB VRAM or larger.

Early abliterated models sometimes trade safety for slight degradation in nuanced reasoning on complex tasks, though newer techniques have narrowed that gap. Whether this release holds up across vision-language workflows or is best suited for text-heavy use cases where the multimodal component is secondary remains to be tested by the community. Benchmark numbers and sample outputs from early users will clarify the real-world trade-offs and whether the abliteration preserves the model's instruction-following quality on uncensored tasks.