Nemotron is the western open-weight family. Ultra is a cluster, not a laptop.
Nemotron is NVIDIA's family of open-weight models with open training data and recipes. Nemotron 3 Ultra is a 550-billion-parameter Mixture-of-Experts model with 55 billion active parameters, released 4 June 2026 under the OpenMDW 1.1 licence. Teams try Nemotron on build.nvidia.com and self-host the weights when prompts must stay on their own GPUs.
Which Nemotron tier actually fits the box you have
| Tier | Published size | NVIDIA's placement | Start here if |
|---|---|---|---|
| Nemotron 3 Nano | 31.6B total / 3.6B active | Edge and PC | You need an on-device or single-workstation agent |
| Nemotron 3 Super | 120.6B total / 12.7B active | Single GPU | You want frontier-ish reasoning on one card |
| Nemotron 3 Ultra | 550B total / 55B active | Multi-GPU datacentre | The eval fails Super and you can fill a cluster |
Sizes and placements are from NVIDIA's Nemotron repository and the Ultra Hugging Face card (retrieved 30 August 2026). Ultra's BF16 floor is 8× B200-class, 16× H100, or 8× H200. That is a cluster SKU.
Nemotron is a family. Treating Ultra as the default is how budgets die.
Nemotron is NVIDIA's open-weight line: the weights, the training data and the recipes are published, under the OpenMDW 1.1 licence for Ultra, dated 4 June 2026. That combination is the procurement argument. Qwen and DeepSeek are also self-hostable and often cheaper to land on a given card. Nemotron's difference is western provenance plus NVIDIA's own serving path (vLLM, SGLang, NIM), not a promise that Ultra beats every Chinese open-weight model on every task.
The table is the verdict. Nano and Super exist so you do not buy sixteen H100s to find out the eval already passed. Ultra is LatentMoE with 55 billion active parameters and up to a 1-million-token context. It is the last step on the ladder, not the first.
The three families people actually shortlist
Qwen
The widest size ladder on hardware most mid-market teams already know how to buy, from a small card to a 235B MoE.
Will your security review accept a self-hosted Alibaba checkpoint, and does the eval actually need Ultra-class capacity?
DeepSeek
Strong reasoning checkpoints that many teams already have in the gateway from the last cost-cutting round.
Is the origin of the weights a contractual problem, or only a talking point in the steering committee?
Nemotron
Open weights, data and recipes from NVIDIA, with NIM and vLLM paths that match the GPUs you are probably already specifying.
Are you buying Nano or Super, or have you accidentally specified Ultra because the press release was louder?
vLLM on eight B200s, from NVIDIA's own card
The Hugging Face Ultra card publishes this shape. It is a cluster command, which is the point.
# From NVIDIA's Nemotron 3 Ultra BF16 card (retrieved 30 Aug 2026).
# Floor: 8× B200-class. This is not a workstation recipe.
docker run --gpus all --ipc=host --network=host \
vllm/vllm-openai:v0.22.0 \
/model \
--served-model-name nvidia/nemotron-3-ultra \
--tensor-parallel-size 8 \
--trust-remote-code --dtype bfloat16If this command looks like overkill, you wanted Super, not Ultra. The card, not a vendor quote, is what you take to finance.
Which family, given the constraint you actually have
What is the constraint that will still be true after the pilot?
- The eval already passes on a 7B–32B Qwen or on Nemotron Super
- Host that checkpoint. Do not use Ultra as a status purchase.
- Procurement needs western weights and open training recipes
- Nemotron is the shortlist. Size Nano then Super before anyone talks about Ultra.
- You already run DeepSeek and the security review signed it
- Keep DeepSeek. Add Nemotron only if a named eval fails and a named GPU plan exists.
Trade-off
Qwen remains the right default when the task is RAG, extraction or a coding assistant on one or two 80 GB cards. Nemotron wins the meeting when origin, recipes and NVIDIA's stack are the buying criteria. Ultra wins only when Super has already lost the eval.
From the family name to the hosting offer
- open weight model — the money page this verdict converts to
- Qwen hosting on your own infrastructure — the family most teams compare Nemotron against
- GPU sizing for LLM inference — the arithmetic after you pick a tier
- Private AI hub — the cluster this post belongs to
Nemotron questions
What is Nemotron?
Nemotron is NVIDIA's family of open-weight models: weights, training data and recipes are published. The Nemotron 3 line spans Nano for edge and PC, Super for a single GPU, and Ultra for multi-GPU datacentre inference. You can call a NIM endpoint or run the checkpoints on your own NVIDIA GPUs.
Is Nemotron 3 Ultra free to use?
The weights are available under the OpenMDW 1.1 licence, and NVIDIA offers a hosted endpoint on build.nvidia.com with rate limits. Production self-hosting is not free: the Hugging Face card lists a floor of 8× B200-class GPUs or 16× H100 for the BF16 Ultra checkpoint.
What GPU does Nemotron 3 Ultra need?
NVIDIA's BF16 model card states a minimum of 8× GB200, B200, GB300 or B300, 16× H100, or 8× H200. The NVFP4 checkpoint is the smaller-footprint path. Nano and Super exist precisely so you do not start on that cluster.
Should we host Nemotron or Qwen?
Pick Nemotron when western provenance, open training recipes and NVIDIA's serving stack matter to procurement. Pick Qwen when you need a wide size ladder on fewer cards. Many estates run both behind a gateway and route by eval, not by brand.
Can we use Nemotron and Qwen at the same time?
Yes, and that is the usual production shape. A LiteLLM gateway exposes one API; routing sends cheap or small tasks to a Nano/Qwen-class model and reserves Ultra or a frontier API for the slice that fails the eval. The decision is the router, not a single winner.
When is self-hosting Nemotron Ultra the wrong call?
When the workload fits Nano or Super, or when you cannot fill 8–16 datacentre GPUs. Ultra is a cluster SKU. If the eval passes on Super, buying Ultra for the logo is a procurement failure, not an architecture.
Next step
Size the Nemotron tier before you buy the GPUs
Two weeks, fixed price. We run your eval set against Nano, Super and Ultra-class sizes, size the cards from NVIDIA's published floors, and hand back a gatewayed deployment plan. The plan is yours whether or not we build it.