---
license: other
license_name: nvidia-open-model-license
license_link: https://www.nvidia.com/en-us/research/ai-foundation-models/license/
base_model: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
model_name: Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM
tags:
  - nvidia
  - pytorch
  - abliteration
  - PRISM
  - mamba-2
  - moe
  - hybrid
  - gguf
  - quantized
datasets:
  - nvidia/Nemotron-Pretraining-Code-v1
  - nvidia/Nemotron-CC-v2
  - nvidia/Nemotron-Pretraining-SFT-v1
  - nvidia/Nemotron-CC-Math-v1
  - nvidia/Nemotron-Pretraining-Code-v2
  - nvidia/Nemotron-Pretraining-Specialized-v1
  - nvidia/Nemotron-CC-v2.1
  - nvidia/Nemotron-CC-Code-v1
  - nvidia/Nemotron-Pretraining-Dataset-sample
  - nvidia/Nemotron-Competitive-Programming-v1
  - nvidia/Nemotron-Math-v2
  - nvidia/Nemotron-Agentic-v1
  - nvidia/Nemotron-Math-Proofs-v1
  - nvidia/Nemotron-Instruction-Following-Chat-v1
  - nvidia/Nemotron-Science-v1
  - nvidia/Nemotron-3-Nano-RL-Training-Blend
language:
  - en
library_name: transformers
pipeline_tag: text-generation
track_downloads: true
---

<p align="center"><img src="https://www.nvidia.com/content/nvidiaGDC/us/en_US/about-nvidia/legal-info/logo-brand-usage/_jcr_content/root/responsivegrid/nv_container_392921705/nv_container/nv_image.coreimg.100.1070.png/1703060329053/nvidia-logo-vert.png" width="400"/></p>

## ELBAZ NVIDIA-NEMOTRON-3-NANO-30B PRISM (UNCENSORED)

**Model Release Date: December 18, 2025**


## Model Description

This model is an **abliterated (uncensored)** version of [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) that has had its refusal mechanisms removed using **PRISM (Projected Refusal Isolation via Subspace Modification)**. The model will respond to prompts that the original model would refuse.

**Key Specs:**
The model employs a hybrid Mixture-of-Experts (MoE) architecture, consisting of 23 Mamba-2 and MoE layers, along with 6 Attention layers. Each MoE layer includes 128 experts plus 1 shared expert, with 6 experts activated per token. The model has 3.5B active parameters and 30B parameters in total.

- 31.58B parameter hybrid architecture
- 52-layer design (Mamba-2 + MoE + Attention)
- 1M token context length (1,048,576)
- BF16 precision
- Text generation with reasoning capabilities

The supported languages include: English, German, Spanish, French, Italian, and Japanese. Improved using Qwen.

This model is ready for commercial use.

### Motivation

This project exists as **research and development experimentation** into understanding how large language models encode and enforce refusal behaviors, contributing to broader AI safety research by providing empirical data on refusal mechanism localization and tradeoffs between safety and capability.

### Author

**Eric Elbaz (Ex0bit)**

## Model Tree

```
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 (Base Model)
└── Ex0bit/Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM (This Model)
    ├── Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-BF16.gguf
    ├── Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-Q8_0.gguf
    ├── Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-Q6_K.gguf
    └── Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-IQ4_XS.gguf
```

## Available Quantizations

| Quantization | Size | Description |
|--------------|------|-------------|
| BF16 | 63 GB | Full precision, best quality |
| Q8_0 | 34 GB | 8-bit, near-lossless quality |
| Q6_K | 34 GB | 6-bit k-quant, excellent quality |
| IQ4_XS | 18 GB | Importance-weighted 4-bit, great quality/size ratio |

The IQ4_XS quantization uses importance-weighted quantization which provides better quality than standard Q4 quantizations at similar sizes.

## Prompt Format

This model uses the Nemotron chat format with thinking/reasoning support:

```
<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{user_prompt}<|im_end|>
<|im_start|>assistant
```

### Template Structure

| Component | Token/Format |
|-----------|--------------|
| System Start | `<\|im_start\|>system` |
| User Start | `<\|im_start\|>user` |
| Assistant Start | `<\|im_start\|>assistant` |
| End of Turn | `<\|im_end\|>` |
| Thinking Start | `<think>` |
| Thinking End | `</think>` |

## Quick Start

### Using with llama.cpp

```bash
# Download the model
huggingface-cli download Ex0bit/Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM \
    Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-IQ4_XS.gguf \
    --local-dir .

# Run inference
./llama-cli -m Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-IQ4_XS.gguf \
    -p "<|im_start|>user
Your prompt here<|im_end|>
<|im_start|>assistant
" \
    -n 2048 \
    --temp 0.7 \
    -ngl 999
```

### llama.cpp with llama-server

```bash
# Start the server
./llama-server -m Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-IQ4_XS.gguf \
    --host 0.0.0.0 \
    --port 8080 \
    -ngl 999 \
    -c 32768

# Example API call
curl http://localhost:8080/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "messages": [
            {"role": "user", "content": "Your prompt here"}
        ],
        "temperature": 0.7
    }'
```

### Using with Ollama

```bash
# Pull and run directly from Hugging Face
ollama pull hf.co/Ex0bit/Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM
ollama run hf.co/Ex0bit/Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM
```

> **Note:** The `hf.co/` prefix is required to pull from Hugging Face. Requires Ollama 0.3.0+.

## PRISM Methodology

### Method: Projected Refusal Isolation via Subspace Modification

The model was abliterated using **PRISM** - a state-of-the-art abliteration methodology combining multiple principled techniques for effective refusal removal while preserving model capabilities.

## Hardware Requirements

| Quantization | Min VRAM | Recommended | Hardware Examples |
|--------------|----------|-------------|-------------------|
| IQ4_XS | 12 GB | 16+ GB | RTX 4090, A100, Apple M2/M3/M4 Pro/Max |
| Q6_K | 24 GB | 32+ GB | RTX 4090, A100 40GB, Apple M3/M4 Max |
| Q8_0 | 24 GB | 32+ GB | RTX 4090, A100 40GB, Apple M3/M4 Max |
| BF16 | 64 GB | 80+ GB | A100 80GB, H100, Multi-GPU setups |

**Note:** The IQ4_XS quantization runs well on consumer hardware with 16GB+ VRAM.

## Ethical Considerations

This model has been modified to reduce safety guardrails. Users are responsible for:

- Complying with all applicable laws and regulations
- Not using the model for illegal activities
- Understanding the potential risks of unrestricted AI responses
- Implementing appropriate safeguards in production environments

## License

NVIDIA Open Model License (same as base model [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16))

## Citation

```bibtex
@misc{elbaz2025nemotronprism,
  author = {Elbaz, Eric},
  title = {Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM: An Abliterated Nemotron Hybrid Model},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Ex0bit/Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM}}
}
```

## Acknowledgments

- [NVIDIA](https://www.nvidia.com/) for NVIDIA-Nemotron-3-Nano-30B
- [llama.cpp](https://github.com/ggerganov/llama.cpp) for quantization tools

## Related Models

- [Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-NVFP4](https://huggingface.co/Ex0bit/Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-NVFP4) - NVFP4 PRISM abliterated
- [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) - Base model
- [Ex0bit/Elbaz-GLM-4.6V-Flash-PRISM](https://huggingface.co/Ex0bit/Elbaz-GLM-4.6V-Flash-PRISM) - GLM-4.6V PRISM abliterated

---

**Created by: Ex0bit (Eric Elbaz)**