New B200 spot capacity is live in US East from $1.69 per GPU-hour. See availability →
Deploy a GPU
NVIDIA · Mixture of experts · released June 2026

Nemotron 3 Ultra GPU requirements: 280 GB at 4-bit, 1.12 TB at BF16

Nemotron 3 Ultra needs 1.12 TB of GPU memory for its weights at BF16, 561 GB at FP8 and 280 GB at 4-bit, plus 0.4 GB of KV cache for a 32K context. The cheapest node that runs it here is 4× A100 PCIe 80GB at 4-bit, for $0.36 an hour on spot.

Architecture read from the model's config.json on Hugging Face, 26 October 2026 · prices from the live catalogue · all 47 GPUs

Cheapest node · 32K context per node-hour, spot
4-bit4× A100 PCIe 80GB · 281 GB needed $0.36$0.76 on-demand
FP88× RTX PRO 6000 Blackwell · 561 GB needed $1.52$3.12 on-demand
BF168× MI300X · 1.12 TB needed $7.92$11.60 on-demand
Deploy 4× A100 PCIe All models

Weights plus a 32K 16-bit KV cache, with 10% of the node's memory kept free. Billed per minute.

Parameters
550B55B active per token
Weights, 4-bit
280 GBBF16: 1.12 TB
KV cache, 32K
0.4 GBOne sequence, 16-bit cache
Context window
256KNative
Memory by format

How much VRAM Nemotron 3 Ultra needs, format by format.

Weights are parameters × bytes per parameter. Each context column adds the KV cache of one sequence of that length.

GPUs by memory
Format Weights + 8Kcontext+ 32Kcontext+ 128Kcontext+ 256Kcontext Cheapest node, 32K
BF16Published weights, fine-tuning 1.12 TB 1.12 TB1.12 TB1.12 TB1.12 TB 8× MI300X · $7.92/h
FP8Hopper, Ada, Blackwell, MI300 561 GB 561 GB561 GB562 GB564 GB 8× RTX PRO 6000 Blackwell · $1.52/h
4-bitAWQ, GPTQ, GGUF Q4 280 GB 280 GB281 GB282 GB283 GB 4× A100 PCIe 80GB · $0.36/h
Arithmetic on published facts, not a benchmark.

The serving engine adds its own buffers, and an 8-bit KV cache halves the context columns. We keep 10% of a node's memory free when we say a model fits. Speed is a separate question: it follows memory bandwidth and, for a mixture of experts, the active parameters.

Where it runs

GPUs that run Nemotron 3 Ultra, cheapest first.

For each format, the smallest node of each GPU model that holds the weights and a 32K context, sorted by spot price. One row per GPU model; larger nodes of the same GPU work too.

All prices
Format GPU GPU memory Spot On-demand 730 h spot
4-bit 4× A100 PCIe 80GBAmpere · 281 GB needed 320 GB $0.36 $0.76 $262.80 Deploy
4-bit 8× A40Ampere · 281 GB needed 384 GB $0.48 $0.72 $350.40 Deploy
4-bit 8× A100 40GBAmpere · 281 GB needed 320 GB $0.72 $1.36 $525.60 Deploy
4-bit 8× L40Ada Lovelace · 281 GB needed 384 GB $0.72 $1.52 $525.60 Deploy
FP8 8× RTX PRO 6000 BlackwellBlackwell · 561 GB needed 768 GB $1.52 $3.12 $1,109.60 Deploy
FP8 8× H100 PCIeHopper · 561 GB needed 640 GB $3.60 $5.52 $2,628.00 Deploy
FP8 8× H100 NVLHopper · 561 GB needed 752 GB $3.92 $6.32 $2,861.60 Deploy
FP8 4× MI300XCDNA 3 · 561 GB needed 768 GB $3.96 $5.80 $2,890.80 Deploy
BF16 8× MI300XCDNA 3 · 1.12 TB needed 1536 GB $7.92 $11.60 $5,781.60 Deploy
BF16 8× MI325XCDNA 3 · 1.12 TB needed 2048 GB $8.72 $12.72 $6,365.60 Deploy
BF16 8× B200 SXMBlackwell · 1.12 TB needed 1440 GB $13.52 $19.92 $9,869.60 Deploy

Spot capacity can be reclaimed with a 2-minute notice; for a live endpoint, use on-demand. How spot works · serving a language model.

Context and the KV cache

What a long context costs in memory.

Hybrid Mamba-Transformer: 12 attention layers (2 KV heads of 128) among 108 blocks; the Mamba layers keep a fixed-size state.

Context KV cache, 1 sequence 8 sequences
8K tokens 0.1 GB 0.81 GB
32K tokens 0.4 GB 3.22 GB
128K tokens 1.61 GB 12.9 GB
256K tokens 3.22 GB 25.8 GB

Serving several users

KV cache

Every concurrent request keeps its own cache, so memory for context grows with users × context length. A server batching 8 requests of 32K needs eight times the single-sequence figure — 3.22 GB here.

  • Recommended node4× A100 PCIe 80GB
  • Weights at 4-bit280 GB
  • Context that fits, 1 sequence256K

Paged attention (vLLM, SGLang) allocates the cache in blocks as requests grow, so a node holds more short requests than the worst case suggests.

Model facts

Nemotron 3 Ultra, as published.

Developer
NVIDIALargest Nemotron 3, hybrid Mamba-Transformer
Parameters
550B total, 55B active per token561B in the published checkpoint (embedding tables, prediction or vision modules included); memory figures use this count
Architecture
Mixture of experts, 108 layersHybrid Mamba-Transformer: 12 attention layers (2 KV heads of 128) among 108 blocks; the Mamba layers keep a fixed-size state
Context window
262,144 tokensAs set in the published configuration
Published format
BF16, 1.12 TB
Licence
Custom licence (see the model card)
Released
3 June 2026
Source
nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16config.json read 26 October 2026

Serve it with vLLM

template: vLLM

On 4× A100 PCIe 80GB, from the instance terminal:

terminal
$ vllm serve nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 \
    --tensor-parallel-size 4 \
    --max-model-len 32768

Very recent architectures need a recent vLLM; the model card gives the minimum version and any extra flags.

FAQ

Nemotron 3 Ultra on a GPU — the questions that come up

How much VRAM does Nemotron 3 Ultra need?

For the weights alone: 1.12 TB at BF16, 561 GB at FP8 and 280 GB at 4-bit. Add 0.4 GB of KV cache for one 32K-token context, or 3.22 GB at the full 256K window, and keep about 10% of the memory free for activations.

Can Nemotron 3 Ultra run on a single GPU?

One RTX 4090 (24 GB): no. One RTX 5090 (32 GB): no. One H100 SXM (80 GB): no. One H200 SXM (141 GB): no. One B200 SXM (180 GB): no. Those answers assume a 32K context and the widest format that still fits.

What is the cheapest way to run Nemotron 3 Ultra?

Here, 4× A100 PCIe 80GB at 4-bit: $0.36 per hour on spot and $0.76 on-demand, billed per minute — $262.80 for a 730-hour month on spot. That node has 320 GB of GPU memory for the 281 GB this setup needs.

Does Nemotron 3 Ultra need less memory because only 55B parameters are active?

No. A mixture-of-experts model routes each token through a few experts, but every expert has to be resident in GPU memory, so the memory bill follows the 550B total. What the 55B active parameters change is the work per token: generation reads far less of the weights each step than a dense model of the same size would, which is why it is fast for its size.

Which format should I run it in?

BF16 keeps the published weights untouched and is the format for fine-tuning. FP8 halves the memory and runs in hardware on Hopper, Ada, Blackwell and MI300-class GPUs. 4-bit quantisation (AWQ, GPTQ, GGUF) quarters it, at a quality cost that is larger on small models than on big ones; check it against your own evaluation before relying on it.

How do I deploy it?

Pick the node in the table above and launch it with the vLLM template; the model downloads from Hugging Face (nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16) on first start. Spread the weights over the node with --tensor-parallel-size equal to the GPU count, and set --max-model-len to the context you need. New architectures need a recent vLLM: check the model card for the minimum version.

Sizing another model? GPU requirements for 64 open models · serving a language model · fine-tuning.

Get started

Nemotron 3 Ultra on 4× A100 PCIe 80GB for $0.36 an hour.

Launch the node with the vLLM template, point it at nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16, stop it when you are done. Prepaid, billed by the minute.

Deploy 4× A100 PCIe All open models

Billed per minute from the moment the instance is reachable. Minimum credit $30, no subscription.