LLM

What GPU do I need to run nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16?

A 550B (MoE) language model for chat and agentic tool use. 560.5B parameters, published in BF16. View on Hugging Face

560.5B
Parameters
BF16
Native precision
256K tokens (262,144)
Context length
NVIDIA OpenMDW-1.1
License
Text
Modality
NVIDIA
Organization
Mixture-of-experts: 22 of 512 experts active per token (~55B active parameters, per the model's own "A55B" name)
Active parameters (MoE)

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 is published by nvidia on Hugging Face, with 287,095 downloads and 337 likes to date. It's a NemotronHForCausalLM model built for text-generation, published natively in BF16.

What NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 is

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 is a 550B-parameter mixture-of-experts language model published by NVIDIA on Hugging Face, part of the Nemotron 3 family. It is released under NVIDIA's own OpenMDW-1.1 license.

License note: NVIDIA's own open model license; read the linked terms before commercial use. Facts in this section are sourced from NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16's Hugging Face model card, not benchmarked by Aquanode.

What it's used for

  • Chat assistants
  • Agentic tool use and function calling
  • Long-context reasoning

VRAM required & cheapest live GPU fit

Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.

PrecisionWeight sizeRequired VRAMCheapest live fitGPUs neededEst. $/hr (full fit)
BF161044.1 GB1252.9 GBAMD MI300X7$18.40/hr
FP8 (quantized)522.0 GB626.4 GBRTX PRO 60007$9.63/hr
INT4 (quantized)261.0 GB313.2 GBRTX A60007$2.54/hr

A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.

INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.

Cheapest way to run NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 at its published (BF16) precision: 7× AMD MI300X, at $2.63/hr per GPU ($18.40/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16: KV cache by context length

The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.

KV-cache size for NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 is not published for this architecture, so no figure is shown.

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16: common questions

Can NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 run on a single GPU?

No. At BF16 it needs 1252.9 GB of VRAM, more than a single 32 GB desktop card holds. The cheapest capable card in the live feed is a 192.0 GB AMD MI300X, and it takes 7 of them.

How many GPUs do I need to run NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16?

7 at BF16. It needs 1252.9 GB of VRAM and the cheapest capable live offer is a 192.0 GB AMD MI300X, so 7 of them come to $18.40/hr in total.

Does quantizing NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 lower the GPU bill?

Yes. At BF16 the cheapest live fit is 7 AMD MI300X cards at $18.40/hr. At INT4 (quantized) it drops to 7 RTX A6000 cards at $2.54/hr, provided a quantized checkpoint exists for it.

How to run NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16

Run NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 with vLLM

From nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16's own model card (8x GPU single-node recipe, trimmed).

docker run -d --name nemotron-ultra-vllm \
  --gpus all \
  --ipc=host \
  --network=host \
  -v $MODEL_CKPT:/model:ro \
  vllm/vllm-openai:v0.22.0 \
  /model \
  --host 0.0.0.0 \
  --port 8000 \
  --served-model-name nvidia/nemotron-3-ultra \
  --trust-remote-code \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --dtype bfloat16 \
  --max-model-len 262144

Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16/raw/main/README.md

Deploy NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 on Aquanode

Aquanode has no one-click deploy template for NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16; you install the inference engine yourself with the commands below. Aquanode sells GPU pods billed per second, not a hosted inference API.

  1. Launch a bare GPU pod sized to the requirement above (7× AMD MI300X or larger).
  2. Open a terminal on the pod, or save one of the commands above as a startup script so it runs automatically the first time the pod boots.
  3. Run the command and connect to the resulting endpoint.
Launch a GPU pod

Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.

More Nemotron 3 models

All 19 Nemotron 3 models: VRAM and GPU requirements

Lineage

Quantized and fine-tuned versions

Alternatives at this size

Other models for text-generation within about a third of NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16's 560.5B parameters, from other model lines.

More on NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16

Related reading: AMD MI300X pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.

Your own GPU cloud accounts

One console for the GPU clouds you already rent from. Your keys stay on your machine.

$20 a month, flat

Connect your accounts

One console for the GPU clouds you already use.

Connect your accounts for $20 a month, flat. Your provider bills the GPU.

© 2026 Aquanode. All rights reserved.

All trademarks, logos and brand names are the property of their respective owners.