Nemotron-Terminal-32B NVFP4

NVFP4 quantization of nvidia/Nemotron-Terminal-32B
Produced by self-quantization on an NVIDIA DGX Spark (GB10 Blackwell) using NVIDIA Model Optimizer.

Model Description

Nemotron-Terminal-32B is NVIDIA's official Qwen3-32B-based model tuned for autonomous terminal tasks, bash scripting, and agentic workflows.

This repo contains a Blackwell-native NVFP4 quantized checkpoint (~20GB), reducing the original ~64GB BF16 footprint by ~70% while preserving quality.

Quantization Details

Property	Value
Base model	nvidia/Nemotron-Terminal-32B
Source precision	BF16
Quantization	NVFP4 (attention: BF16, experts: NVFP4)
Tool	NVIDIA Model Optimizer (nvidia-modelopt 0.41.0)
Calibration	512 samples, synthetic random
Hardware	NVIDIA GB10 Blackwell (DGX Spark)
Output size	~20GB (vs ~64GB BF16)

Usage with vLLM

vllm serve kleinpanic93/Nemotron-Terminal-32B-NVFP4 \
  --quantization modelopt_fp4 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.72 \
  --trust-remote-code \
  --enable-chunked-prefill \
  --enable-prefix-caching

Note: NVFP4 kernels require NVIDIA Blackwell (GB10/GB200) or newer hardware. Does not run on Hopper (H100) or older.

Provenance

{
  "source_model": "nvidia/Nemotron-Terminal-32B",
  "quantization": "NVFP4",
  "tool": "nvidia-modelopt",
  "hardware": "NVIDIA GB10 (Blackwell)",
  "attention_dtype": "bfloat16"
}