Arrakis / Models / NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 GPU requirements
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 has 561B parameters. To serve it, plan for about 1,345 GB of GPU memory in 16-bit, 673 GB in 8-bit or 336 GB in 4-bit, including about 20% for the KV cache and activations.
It is a mixture of experts: 512 experts, 22 of them active for each token. All of them stay in GPU memory, so plan memory for all 561B parameters.
What it takes to serve it
GPU memory to run it for inference, with the smallest GPU setup that holds it and the lowest price on Arrakis right now.
| Precision | Memory and the smallest setup |
|---|---|
| 16-bit | 1,345 GB: 8 × B200, from $54.32 per hour for all 8 |
| 8-bit | 673 GB: 4 × B200, from $27.16 per hour for all 4 |
| 4-bit | 336 GB: 2 × B200, from $13.58 per hour for all 2 |
What it takes to fine-tune it
QLoRA trains adapters on a 4-bit frozen model, LoRA on a 16-bit one, and full fine-tuning updates every weight. Figures are before activations, which grow with batch size and sequence length.
| Method | Memory and the smallest setup |
|---|---|
| QLoRA | 392 GB: 2 × B300, from $9.20 per hour for all 2 |
| LoRA | 1,345 GB: 8 × B200, from $54.32 per hour for all 8 |
| Full fine-tuning | 8,968 GB: more than one 8-GPU machine |
About the model
From its card and files on Hugging Face. Memory figures here follow the GPU memory guide.
- Parameters
- 561B
- Active for each token
- 55B
- Experts
- 512
- Experts for each token
- 22
- Context window (tokens)
- 262,144
- Published in
- BF16
- License
- other
- Released
- 2026-06-03
Deploy it on Arrakis
Open Models in the console, choose vLLM, SGLang or llama.cpp, and rent a GPU it fits on. You get an OpenAI-compatible endpoint with its own API key, and root SSH to the machine. Or run it yourself on any machine with enough GPU memory:
vllm serve nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 --tensor-parallel-size 8Questions
Live prices for every GPU are on the GPU prices page.