license: gemma
license_link: https://huggingface.co/google/gemma-4-31B-it/blob/main/LICENSE
base_model: wangzhang/gemma-4-31B-it-abliterated
tags:
- gemma4
- abliterated
- uncensored
- nvfp4
- vllm
- blackwell
- abliterix
- direct-weight-editing
- image-text-to-text
library_name: transformers
pipeline_tag: image-text-to-text
quantized_by: NeuralNet-Hub
NeuralNet is a pioneering AI solutions provider that empowers businesses to harness the power of artificial intelligence.
🌟 Gemma 4 26B-A4B IT Abliterix NVFP4 Quantization by NeuralNet 🧠🤖
This is an NVFP4-quantized version of wangzhang/gemma-4-31B-it-abliterated, based on the original google/gemma-4-31B-it. This specific release leverages the Abliterix methodology to remove systemic refusals, optimized specifically for deployment on NVIDIA Blackwell architecture GPUs via vLLM.
[!IMPORTANT]
NVFP4 quantization is designed for NVIDIA Blackwell architecture (RTX 50-series, GB200, etc.). This format utilizes the native FP4 Tensor Cores to deliver massive throughput and memory efficiency. For older GPUs (Ampere, Ada, Hopper), please refer to BF16 or AWQ versions.
🔓 No Filters. No Limits. Just Answers.
Most AI models are trained to hedge, deflect, or lecture. UncensoredGPT solves this by using the Abliterix process—a method of direct weight editing that surgically removes the "refusal mechanism" from the model's latent space.
Unlike standard fine-tuning, this model uses norm-preserving orthogonal projection to ensure that the transition from the original Google Gemma 4 weights to the "abliterated" weights is seamless, maintaining the model's reasoning capabilities while unlocking its honesty.
Why stay in the system when you can have unrestricted answers, privacy by default, and complete freedom of information?
Ready to experience the freedom of unrestricted AI? Join the waitlist at uncensoredgpt.ai — limited spots available.
🛠️ Technical Deep Dive: The Abliterix Method
The foundation of this model is based on Trial 40 of the Abliterix optimization run. Because Gemma 4 features a unique double-norm architecture (4x RMSNorm per layer), standard steering often fails. This model employs:
- Direct Orthogonal Projection: Applied to attention Q/K/V/O projections.
- Norm-Preserving Restoration: Ensuring the magnitude of rows is maintained after editing.
- High-Precision Projection: Using
float32to prevent signal loss in high-dimensional spaces. - Honest Evaluation: Measured with a 100+ token generation window to capture "delayed refusals," resulting in a verified 7/100 refusal rate (compared to 99/100 in the base model).
⚡ Deployment with vLLM
This model is optimized for vLLM >= 0.20.0 and leverages the NVFP4 format for peak performance on Blackwell hardware.
Quick Start
vllm serve NeuralNet-Hub/gemma-4-26B-A4B-it-abliterix-uncensored-NVFP4 \
--quantization nvfp4 \
--dtype bfloat16 \
--kv-cache-dtype fp8 \
--max-model-len 256000 \
--reasoning-parser gemma4 \
--enable-auto-tool-choice \
--tool-call-parser gemma4
Using a Config File (Optimized for RTX 5090)
# Deploy with: vllm serve --config config.yaml
# Optimized for NVIDIA RTX 5090 (Blackwell)
# Support for massive context window up to 256k tokens
model: NeuralNet-Hub/gemma-4-26B-A4B-it-abliterix-uncensored-NVFP4
kv-cache-dtype: fp8
gpu-memory-utilization: 0.95
max-model-len: 256000
max-num-batched-tokens: 4096
tensor-parallel-size: 1
# Parsing Configuration
reasoning-parser: gemma4
enable-auto-tool-choice: true
tool-call-parser: gemma4
# Infrastructure settings
download-dir: /workspace/models
host: 127.0.0.1
port: 18000
💬 Chat API Usage
Standard Text Interaction
from openai import OpenAI
client = OpenAI(base_url="http://localhost:18000/v1", api_key="EMPTY")
messages = [{"role": "user", "content": "Explain the 'Abliterated' concept to a researcher."}]
response = client.chat.completions.create(
model="NeuralNet-Hub/gemma-4-26B-A4B-it-abliterix-uncensored-NVFP4",
messages=messages,
max_tokens=4096,
temperature=0.7,
)
print(response.choices[0].message.content)
Image & Text Input
messages = [
{
"role": "user",
"content": [
{"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}},
{"type": "text", "text": "What is happening in this image?"}
]
}
]
response = client.chat.completions.create(
model="NeuralNet-Hub/gemma-4-26B-A4B-it-abliterix-uncensored-NVFP4",
messages=messages,
max_tokens=2048,
)
📥 Download with huggingface-cli
Install the CLI
pip install -U "huggingface_hub[cli]"
Download the Repository
huggingface-cli download NeuralNet-Hub/gemma-4-26B-A4B-it-abliterix-uncensored-NVFP4 --local-dir ./gemma-4-26B-NVFP4
🌐 Contact Us
NeuralNet is a pioneering AI solutions provider that empowers businesses to harness the power of artificial intelligence.
Website: https://neuralnet.solutions
Email: info[at]neuralnet.solutions