Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM-NVFP4 (UNCENSORED)

NVFP4 Quantized Version for NVIDIA Blackwell GPUs

Model Description

This is the NVFP4 (FP4) quantized version of Ex0bit/Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM, created using NVIDIA TensorRT Model Optimizer.

Model Size: ~18 GB (4-bit weights)

Requirements

NVFP4 format is optimized for:

• NVIDIA Blackwell GPUs (B100, B200, GB200)
• TensorRT-LLM v0.17+

This format is NOT compatible with:

• Consumer GPUs (RTX 3000/4000/5000 series)
• llama.cpp
• Standard PyTorch inference

Usage

With TensorRT-LLM

Deploying NVIDIA Nemotron-3-Nano with TensorRT LLM

Quantization Details

• Method: NVIDIA ModelOpt NVFP4_DEFAULT_CFG
• Calibration: 512 samples from WikiText-2
• Format: HuggingFace checkpoint with quantized weights

Related Models

• Ex0bit/Elbaz-NVIDIA-Nemotron-3-Nano-30B-A3B-PRISM - Parent model (GGUF quantizations available)
• nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 - Original base model

Author

Eric Elbaz (Ex0bit)

License

NVIDIA Open Model License