Local AI hardware
Sizing table
| Model | Checkpoint | Weights (estimate) | Fits on |
|---|---|---|---|
| Nemotron Nano 12B v2 VL | BF16 / FP8 | ~26 GB / ~14 GB | RTX 4090 24 GB, RTX 5090 32 GB |
| Nemotron 3 Nano | FP8 / NVFP4 | ~32 GB / ~18 GB | RTX 5090 32 GB, L40S 48 GB |
| Nemotron 3 Super | NVFP4 / BF16 | ~65–70 GB / ~240 GB | A100 80 GB, H100 80 GB, RTX PRO 6000 96 GB |
| Nemotron 3 Ultra | NVFP4 / BF16 | ~280–300 GB / ~1.1 TB | 4× RTX PRO 6000, DGX B200, clustered DGX Spark |
These numbers are our own arithmetic — roughly 0.55 bytes × parameters for NVFP4 weights plus
activations and KV cache. NVIDIA does not publish a VRAM table for these checkpoints. Measure
before you buy, and treat this page as a starting point rather than a guarantee.
DGX Spark
A GB10 machine with 128 GB of unified memory and NVFP4 throughput in the petaflop class is the natural single-box target for Nemotron 3 Nano and, with a capped context, Super. It is not a single-node target for Ultra.
The appliance runs on it unchanged: same ISO, same UI, same policies. The only difference is that
the local-nemotron pack has a GPU to attach to, so the provider profile can point at localhost.
No GPU at all
Nothing breaks. Use cloud models, keep the policies identical, and move the provider profile to a local endpoint when the hardware arrives. That is the recommended way to evaluate the workflow before spending money.