Uploaded September 2026 | Updated November 2026, 3 days ago
Learn how NVFP4 compares to BF16 in practice, why the smaller checkpoint frees up GPU memory without a big quality hit, and where each format still makes the most sense.
Using Nemotron 3.5 Lightning, which ships as both a BF16 and an NVFP4 checkpoint, we look at memory footprint, throughput, and accuracy, then walk through a notebook that runs the whole comparison on a single H100 so you can reproduce these results.
What Is NVFP4?: youtu.be/UTfg-_EGurw?si=8V9kG1AJi4O4NIFB
Notebook: github.com/patelmiteshn/inference-cookbooks/blob/main/examples/nvfp4_vs_bf16.ipynb
Learn how NVFP4 compares to BF16 in practice, why the smaller checkpoint frees up GPU memory without a big quality hit, and where each format still makes the most sense.
Using Nemotron 3.5 Lightning, which ships as both a BF16 and an NVFP4 checkpoint, we look at memory footprint, throughput, and accuracy, then walk through a notebook that runs the whole comparison on a single H100 so you can reproduce these results.
What Is NVFP4?: youtu.be/UTfg-_EGurw?si=8V9kG1AJi4O4NIFB
Notebook: github.com/patelmiteshn/inference-cookbooks/blob/main/examples/nvfp4_vs_bf16.ipynb










