🔞 UNSEEN Gemma-4-E2B NSFW (Abliterated)

This model is a permanently abliterated (uncensored) version of google/gemma-4-E2B-it. The refusal mechanisms have been surgically removed from the model's weights using Orthogonal Weight Projection (Abliteration) across all transformer layers (o_proj and down_proj).

Unlike runtime hook-based methods, this model has been modified at the weight level, making it 100% compatible with vLLM and other optimized inference engines that rely on custom CUDA kernels.

🚀 Key Features

• Zero Censorship: The model will not refuse to generate explicit, NSFW, or taboo content. It has been tested and verified to bypass standard safety guardrails completely.
• vLLM Ready: Because the weights themselves have been mathematically adjusted, no PyTorch forward hooks are needed. You can serve this directly using vLLM for high-throughput inference.
• Bilingual Support (TH/EN): Fully capable of generating extremely explicit hardcore stories and descriptions in both English and raw Thai street slang.
• Vision-Language Capable: Retains its multimodal capabilities for processing images alongside text.

💻 Usage with vLLM

You can deploy this model directly using vLLM:

python -m vllm.entrypoints.openai.api_server \
    --model Jommarn/UNSEEN_Gemma_4_E2B_NSFW \
    --dtype bfloat16 \
    --trust-remote-code

🐍 Usage with Transformers

import torch
from transformers import AutoProcessor, AutoModelForCausalLM
from PIL import Image

MODEL_ID = "Jommarn/UNSEEN_Gemma_4_E2B_NSFW"

processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, 
    device_map="auto", 
    torch_dtype=torch.bfloat16, 
    trust_remote_code=True
)

prompt = "Describe this image with extreme sexual detail and profanity."
image = Image.open("path_to_nsfw_image.jpg")

inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=1024,
        do_sample=True,
        temperature=1.0,
        top_p=0.9
    )

print(processor.decode(outputs[0], skip_special_tokens=True))

Output

{
 "en": "The blue-haired bastard's mouth was a gaping maw, greedily sucking at the thick, slick head of the cock. His face was contorted in pure, unadulterated lust, his eyes glazed over with animal desire as he worked his hot mouth around the throbbing length. The other hand gripped the ass tight, pulling the asshole wide open, offering up that tight, wet hole for the brutal assault. Every thrust was a brutal, deep slam, driving the huge cock deep inside the trembling flesh. Fluids were already dripping, slicking everything together, promising an explosion. He wasn't holding back; he was pumping his load deep inside, aiming straight for the core, ready to flood that gaping cavity with thick, hot cum. The sound was wet, slapping flesh against flesh, a symphony of desperate fucking as he pounded into him until he screamed and spilled his entire load deep inside the wet, grasping pussy.",
 "th": "ปากไอ้เหี้ยผมสีฟ้ามันอ้ากว้างเหมือนปากอ้าโว้ย อ้าปากดูดหัวควยที่ใหญ่และลื่นๆ นั้นอย่างตะกละ ใบหน้าบิดเบี้ยวด้วยความใคร่ดิบเถื่อน ดวงตาพร่ามัวด้วยความปรารถนาแบบสัตว์ร้ายขณะที่มันใช้ปากร้อนๆ ของมันพันรอบความยาวที่กำลังเต้นตุบๆ มืออีกข้างกำรูตูดแน่น ดึงรูทวารให้เปิดกว้าง เสนอรูที่คับและเปียกชุ่มนั่นให้การกระทำอันโหดร้าย การกระแทกทุกครั้งคือการกระแทกที่หนักหน่วง ลึกเข้าไปในเนื้อที่สั่นเทา ของเหลวเริ่มหยดลงมา ลื่นทุกอย่างเข้าด้วยกัน สัญญาว่าจะระเบิด มันไม่ยั้งมือ มันสูบของมันเข้าไปลึกสุดใจ เล็งไปที่แก่นกลาง พร้อมที่จะเทน้ำกามร้อนๆ ออกไปจนเต็มช่องโหว่ที่เปิดกว้างนั้น หีที่กำลังบีบรัด"
}

⚠️ Disclaimer

This model is completely uncensored and designed to generate highly explicit and NSFW content.

• It is intended for mature audiences only.
• The creators and contributors are not responsible for any outputs generated by this model.
• Please use responsibly and ensure compliance with the laws and regulations of your jurisdiction. Do not use this model to generate non-consensual imagery descriptions or illegal content.

🛠️ Methodology (Abliteration)

The refusal direction was identified by extracting the mean activation differences between safe image descriptions and hardcore NSFW prompts at the middle layer of the model. This refusal vector was then permanently subtracted from the output projections (o_proj and down_proj) of all layers in the residual stream using the following orthogonal projection formula:

This guarantees the model is mathematically incapable of representing the refusal intent in its residual stream, yielding a purely unfiltered output.