How much memory does Nemotron 3 Ultra need in FP16?
Nemotron 3 Ultra has 550 billion parameters. The FP16 format decides how much memory is needed to load them.
How is this footprint calculated?
Nemotron 3 Ultra totals 550 billion parameters. The FP16 format takes 2 byte per parameter (16 bits per parameter, hence 2 bytes). The product gives the weights, to which a 15% runtime margin is added for activations and buffers: 550 billion parameters × 2 bytes (FP16) = 1,100.0 GB of weights; × 1.15 runtime margin = 1,265.0 GB. The attention cache is not calculated on this page. These models’ architectures (compressed latent, linear or sparse attention, Mamba layers) share no per-token formula; it is measured on the machine, with the target context and concurrency [TO BE MEASURED].
What does the format change?
FP16 brings full precision, the quality reference.
| Format | Weights in memory | Compatible platforms |
|---|---|---|
| FP16 | 1,265 GB | 3 |
| FP8 | 632 GB | 6 |
| NVFP4 | 316 GB | 7 |
| Q4 | 379 GB | 7 |
Which platforms qualify?
See the Nemotron 3 Ultra fact sheet.
How much memory for Nemotron 3 Ultra in FP16?
About 1,265 GB for the weights, excluding the attention cache.
Which format should I choose for Nemotron 3 Ultra?
FP16 brings full precision, the quality reference. The most precise format that fits the target machine remains the best choice.
Method and sources
Memory footprints are calculated, not measured: a reading on real hardware may differ depending on the engine and the exact weight format. Every figure below carries its source and the date of its last check.
- Nemotron 3 Ultra: 550 billion parameters, 55 billion active, 262,144-token context, OpenMDW 1.1 licence. Hugging Face repository (config.json, API, model card), checked on 2 September 2026. 550 billion announced; the BF16 repository holds 560 billion elements (1,121 GB). The card claims “up to 1M” context, config.json declares 262,144.
- Bytes per parameter: FP16 2 (16 bits per parameter, hence 2 bytes); FP8 1 (8 bits per parameter (E4M3), hence 1 byte, block scales not counted); NVFP4 0.5 (4 bits per parameter, hence 0.5 byte; 4-bit published weights weigh 0.54 to 0.56 byte per parameter with their scales (DeepSeek V4 Pro 865 GB for 1,599 billion, Kimi K3 1,561 GB for 2,780 billion)); Q4 0.6 (taken here as llama.cpp Q4_K_M, 0.6 byte per parameter including scales; Q4_K_S weighs 0.56, AWQ and GPTQ 0.55). Source: connaissance/faits.yaml, quantifications family, checked on 2026-08-31.
- Runtime margin × 1.15: QDNA operating assumption, not measured: 15% above the weights for activations, buffers and fragmentation.
- “Fits” threshold at 75% of memory: QDNA assumption, not measured, which keeps the remaining quarter for the attention cache and concurrency.