🤖 Huihui-gemma-4-12B-it-abliterated: Google Gemma 4 without safety filters
A version of Google Gemma 4 12B-it with removed safety filters has been released. To eliminate refusals, the abliteration method was used, affecting only layers 23–28. The model supports multimodality and requires approximately 16 GB of VRAM.
🌍 This demonstrates the effectiveness of the abliteration method for quickly removing restrictions without full fine-tuning, confirming the trend toward creating uncensored models for research.
👤 It is now possible to utilize the capabilities of Gemma 4 without intrusive refusals, which is critical for creative tasks, role-playing, and deep technical analysis.
Source 1: https://huggingface.co/huihui-ai/Huihui-gemma-4-12B-it-abliterated