Nemotron 3 Ultra 550B-A55B has 560.5B parameters (55B active per token). In BF16 its weights alone take 1023 GiB; quantized to 4-bit, about 309 GiB. Each 32K-token request adds 0.38 GiB of KV cache. Cheapest way to run it today: 7 × RTX A6000 (48 GB) at $2.81 per hour. It calls tools and has a thinking mode you can switch off.
Read every day from Hugging Face. Split files are added together; vision projectors are left out.
How many GPUs to run Nemotron 3 Ultra 550B-A55B
Fewest GPUs that fit the weights plus one 8K-token request. Data-center GPUs use vLLM (90% of memory usable), the others llama.cpp. Several GPUs split the model with tensor or pipeline parallelism.
GPU
NVFP4 (native)
BF16 / FP16
8-bit
4-bit
GeForce RTX 4090 24 GB
Not on this GPU
64 (several machines)
32 (several machines)
16 (several machines)
GeForce RTX 5090 32 GB
Not on this GPU
48 (several machines)
24 (several machines)
16 (several machines)
RTX PRO 6000 Blackwell 96 GB
Not on this GPU
16 (several machines)
6
4
Mac M5 Max 128 GB 96 GB usable
Not on this GPU
16 (several machines)
6
4
Ryzen AI Max+ 395 128 GB 96 GB usable
Not on this GPU
16 (several machines)
6
4
DGX Spark (GB10) 128 GB
Not on this GPU
16 (several machines)
5
3
NVIDIA H100 80 GB
Not on this GPU
16 (several servers)
8
5
NVIDIA H200 141 GB
Not on this GPU
16 (several servers)
5
3
NVIDIA B200 (HGX) 180 GB
2
7
4
2
8-bit: FP8 with vLLM (INT8 on GPUs without FP8), Q8_0 with llama.cpp. 4-bit: AWQ / GPTQ with vLLM, Q4_K_M with llama.cpp, both counted at the Q4_K_M size. Not on this GPU: the official NVFP4 weights do not run there (llama.cpp needs a GGUF version, and vLLM needs a GPU with FP4 support).
Cheapest way to run Nemotron 3 Ultra 550B-A55B today
For each GPU: the fewest cards that fit Nemotron 3 Ultra 550B-A55B with one 8K-token request, the most faithful weight format at that count, and the cheapest on-demand price (2026-10-30).
On-demand prices from AWS, Google Cloud, Azure, RunPod, Vast.ai, Oracle Cloud, OVHcloud, Scaleway, Verda, Akash and UpCloud, checked every hour. Serving many users needs more KV cache, so more memory: size it in the calculator. Some links are affiliate links: StudioTV may earn a commission, at no extra cost to you.
vLLM and SGLang apply these defaults unless a request sets its own. With llama.cpp, pass them yourself: --temp 1 --top-p 0.95
Reasoning. Thinking mode you can switch on or off. With vLLM or SGLang, set it per request: "chat_template_kwargs": {"enable_thinking": false} turns it off, true on. Reasoning tokens take context and KV cache like any other token: leave room for them.
KV cache: how memory grows with context
Architecture: KV on 12 of 108 layers (GQA, 2 KV × 128) · 48 Mamba-2 layers · 48 LatentMoE layers. Its KV cache is 9.0× smaller than if every layer used full attention, because only 12 of its 108 layers keep a KV cache; the others hold a small fixed-size state. Only 55B parameters are active per token, but all 561B must sit in memory: VRAM depends on the total, speed on the active part.
Context per request
BF16 cache
FP8 cache
8K tokens
0.09 GiB
0.05 GiB
32K tokens
0.38 GiB
0.19 GiB
128K tokens
1.50 GiB
0.75 GiB
256K tokens
3.00 GiB
1.50 GiB
Plus 389 MiB of recurrent state per request, whatever the context.
Nemotron 3 Ultra 550B-A55B VRAM FAQ
How much VRAM does Nemotron 3 Ultra 550B-A55B need?
In BF16 the weights alone take 1023 GiB (1099 GB). In FP8 that is 514 GiB, and about 309 GiB with 4-bit quantization (Q4_K_M); the official NVFP4 checkpoint is 307 GiB. Each request then adds KV cache: 0.09 GiB at 8K tokens and 0.38 GiB at 32K (BF16 cache).
Can Nemotron 3 Ultra 550B-A55B run on a single RTX 4090 (24 GB)?
No. Even in 4-bit it needs 16 RTX 4090s, spread over several machines. In 4-bit it needs 5 H100 GPUs.
Can Nemotron 3 Ultra 550B-A55B run on an RTX 5090 (32 GB)?
No. Even in 4-bit it needs 16 RTX 5090s, spread over several machines.
Can Nemotron 3 Ultra 550B-A55B run on a Mac?
Yes. In 4-bit (Q4_K_M with llama.cpp) it needs a Mac with 512 GB of unified memory, of which macOS gives the GPU about 384 GB, for one request with an 8K-token context. In 8-bit it does not fit on a single Mac.
Can Nemotron 3 Ultra 550B-A55B run on a DGX Spark or a Ryzen AI Max+ 395 (128 GB)?
No. Even in 4-bit it needs 3 DGX Spark units or 4 Ryzen AI Max+ 395 machines (128 GB), linked over the network.
What is the cheapest way to run Nemotron 3 Ultra 550B-A55B?
On 2026-10-30, the cheapest on-demand setup is 7 × RTX A6000 (48 GB) with Q4_K_M weights on Vast.ai, at $2.81 per hour (about $2,049.11 per month). Next: 5 × NVIDIA A100 80GB with 4-bit weights on Vast.ai, at $3.15 per hour. Sized for one 8K-token request; prices are checked every hour.
How many H100 GPUs does Nemotron 3 Ultra 550B-A55B need?
With one request and an 8K-token context: 16 in BF16, 8 in FP8 and 5 in 4-bit (vLLM, 90% of memory usable). Above 8 GPUs, that means several servers. Serving many users at once needs more memory for their KV caches: the calculator sizes that for you.
Does Nemotron 3 Ultra 550B-A55B support tool calling?
Yes, Nemotron 3 Ultra 550B-A55B supports tool (function) calling: it can ask your application to run a function and use the result.
Does Nemotron 3 Ultra 550B-A55B have a thinking mode?
Thinking mode you can switch on or off. With vLLM or SGLang, set it per request with "chat_template_kwargs": {"enable_thinking": false} (or true). Reasoning tokens count toward the context and the KV cache.
What temperature should I use with Nemotron 3 Ultra 550B-A55B?
The official defaults are temperature 1, top-p 0.95.
Nemotron 3 Ultra 550B-A55B VRAM badge
For a model card or a README: the 4-bit size and the smallest GPU it fits on, linked to this page. Data as JSON: /api/models/nemotron-3-ultra.json.
Estimates, not guarantees: computed from the official config.json with the same engine as the LLM VRAM Calculator. Real usage depends on your engine version and settings.
NVIDIA, United States · Released June 4, 2026