NVIDIA Nemotron 3: CUDA-first open AI

15 October 2025 · 8 min · Model Profiles

TL;DR Nemotron 3 Ultra and Super are NVIDIA's open-weight variants: Ultra in the 550B-total / 55B-active class, Super smaller, with 262K to 1M context. They are the CUDA-first option in this survey, heavily optimized for NVIDIA hardware and tooling. The strongest use case is a deployment stack that is already CUDA-centric, where NVIDIA integration saves real engineering. Node-class hardware with excellent acceleration when run on the intended vendor stack.

Most open-model releases answer "what is the best model?" with a parameter count. NVIDIA's Nemotron series answers a different question: "what is the best model for a data centre that already runs on our hardware?" Nemotron 3, with its Ultra and Super open variants, is the vendor-native option in this survey. This post covers what changed, what it takes, and the one condition under which none of the other models matter as much.

What is Nemotron 3?

Nemotron 3 is NVIDIA's open-weight line, with Ultra and Super leading the family [model card](https://huggingface.co/nvidia/Nemotron-3) (P1). Ultra is the larger variant at around 550 billion total parameters with roughly 55 billion active, and Super is a smaller sibling, with contexts reported from 262K up to 1M tokens depending on variant [NVIDIA announcement](https://developer.nvidia.com/blog) (P3). Independent trackers list them as competitive mid-to-large options [BenchLM](https://benchlm.ai) (P2), [Artificial Analysis](https://artificialanalysis.ai) (P2), with the details varying by variant.

What is the CUDA-first argument?

NVIDIA builds these models to run well on NVIDIA hardware, with NVFP4 and FP8 alignment, TensorRT-LLM support, and toolchain integration. If your data centre, servers and operations team are already NVIDIA-native, this saves you engineering weeks that a model from another vendor would spend on adaptation. If your stack is mixed or non-NVIDIA, that optimization is worth much less to you, and permissive alternatives from other labs become the better trade.

How do the variants differ?

Ultra is the heavier generalist for the hardest jobs on the biggest nodes. Super is the lighter sibling for the same vendor-native path at a smaller cost. For most teams the choice is hardware budget: run Ultra on the largest node you have, Super when the workload or the card count says so.

Can you legally use it commercially?

Confirm the exact license for the specific Nemotron 3 variant on the card (P1). The open variants in this series carry NVIDIA's standard terms, which are permissive enough for ordinary internal and commercial use, with the controlling document being the card's LICENSE file. The general framework for reading these terms is in the post on [what open-source AI means after the license shifts](https://dintechnologies.com/blog/what-open-source-ai-means-after-2025-license-shifts).

What hardware does it take?

Everything about Nemotron assumes NVIDIA hardware, and the answer scales accordingly:

• Ultra needs a serious multi-GPU node; Super is the lighter step down.
• Expect the natural companions of the ecosystem: high-memory NVIDIA cards, NVFP4 and FP8 storage, TensorRT-LLM.
• The [quantization guide](https://dintechnologies.com/blog/2026-quantization-moe-efficiency-memory-optimization) matters even more here, because the precision stack is part of the acceleration.

What does it take to run it?

1. Serve with TensorRT-LLM or the NVIDIA-native runtime path; the engine comparison is in [local AI runtimes compared](https://dintechnologies.com/blog/2026-inference-runtimes-local-ai-tooling-maturity).

2. Use NVFP4 or FP8 builds to keep the node targeted to the memory you have.

3. Size context deliberately; a 262K to 1M window is a scheduling and memory decision per request.

4. Keep the deployment inside your network for private workloads, with the same logic as [running open models privately](https://dintechnologies.com/blog/running-open-models-privately).

What does it cost?

The [cost reality check](https://dintechnologies.com/blog/mid-year-cost-reality-check) is written for exactly this conversation: if the hardware is already NVIDIA and already paid, running Nemotron adds the smallest marginal cost while capturing its optimizations. If the hardware is not NVIDIA, the same model loses its advantage and the comparison changes entirely.

When is Nemotron 3 the wrong choice?

• Your stack is not NVIDIA-centric: the vendor optimization is lost, and permissive alternatives score better per euro.
• You need the top marks on open reasoning tables: the flagships are ahead.
• You deploy on consumer or mixed hardware: the targeted optimizations are wasted.

Where this leaves you

If you run an NVIDIA-native data centre, Nemotron 3 is the open model to evaluate first, because its integration cost is the lowest of any option in this survey. If you run anything else, treat it as a mid-to-large open model with a good card and move on. Either way, verify the specific variant's license and context before adoption, and let your hardware decide the rest.