Jommarn/UNSEEN_Gemma_4_26B_NSFW
Jommarn/UNSEEN_Gemma_4_26B_NSFW is a 26 billion parameter Gemma-4-26B-A4B-it based model, permanently modified to be uncensored. Its refusal mechanisms have been surgically removed from the weights using Orthogonal Weight Projection (Abliteration) across all transformer layers, making it fully compatible with vLLM. This model excels at generating highly explicit, NSFW, and taboo content in both English and Thai, while retaining its multimodal vision-language capabilities.
Loading preview...
Overview
Jommarn/UNSEEN_Gemma_4_26B_NSFW is a 26 billion parameter model derived from google/gemma-4-26B-A4B-it. Its core distinction lies in the permanent removal of censorship and refusal mechanisms through a technique called Orthogonal Weight Projection (Abliteration). This modification was applied directly to the model's weights, specifically the o_proj and down_proj layers across all transformers, ensuring a truly uncensored output.
Key Capabilities
- Zero Censorship: Designed to generate explicit, NSFW, and taboo content without refusal, bypassing standard safety guardrails.
- vLLM Compatibility: The weight-level modification ensures 100% compatibility with vLLM and other optimized inference engines, as no PyTorch forward hooks are required.
- Bilingual Support: Capable of generating highly explicit content in both English and Thai street slang.
- Vision-Language: Retains its multimodal capabilities, allowing it to process images alongside text for explicit descriptions.
Methodology
The model's refusal direction was identified by analyzing mean activation differences between safe and NSFW prompts. This refusal vector was then mathematically subtracted from the output projections of all layers, making the model inherently incapable of representing refusal intent. Users are advised to use specific generation parameters (do_sample=True, temperature=1.0) and a "Prefill" technique to guide the model towards desired explicit content.