NVIDIA Nemotron 3 Family of Models: Nano, Super, and Ultra
A practical guide to the three core NVIDIA Nemotron 3 models. Compare Nano, Super, and Ultra, understand the hybrid architecture, see what each model is designed for, and find official research, checkpoints, and deployment resources.
Quick answer: NVIDIA Nemotron 3 is a family of open AI language models built for reasoning and agentic workloads. Nano prioritises efficient inference for specialised tasks, Super is aimed at collaborative agents and demanding high-volume workflows, and Ultra is the largest option for complex reasoning and long-running agent systems. The best choice depends on your task, serving environment, latency target, and available hardware—not parameter count alone.
What is the NVIDIA Nemotron 3 family of models?
NVIDIA Nemotron 3 is a family of open models developed for applications that need more than a short chat response. The models are designed to support reasoning, tool use, conversational tasks, and agentic workflows where an AI system may need to plan, work through several steps, and use tools.
The original core family contains three model sizes: Nemotron 3 Nano, Nemotron 3 Super, and Nemotron 3 Ultra. They share a hybrid model-design direction, but they are not simply three interchangeable versions of the same checkpoint. Each is positioned for a different balance of inference efficiency, capability, and workload scale.
NVIDIA announced the family on December 15, 2025. Since that announcement, separate technical reports and model pages have provided more complete specifications for Super and Ultra. For that reason, this guide uses the later technical reports where available rather than relying only on the early launch announcement.
For the official family overview, see NVIDIA Research: Nemotron 3 Family of Models.
NVIDIA Nemotron 3 Nano vs Super vs Ultra
The simplest way to understand the lineup is to compare its intended workload and the number of parameters each model activates during a forward pass. The values below are taken from NVIDIA's model reports and official model pages; rounded model names and parameter accounting can differ slightly between documents.
| Model | Total parameters | Active parameters | Best-fit workload | Context |
|---|---|---|---|---|
| Nemotron 3 Nano | 31.6B | 3.2B per pass (3.6B including embeddings) | Efficient specialised tasks and sub-agents | Up to 1M tokens |
| Nemotron 3 Super | About 120B | About 12B | Collaborative agents and high-volume reasoning workflows | Up to 1M tokens |
| Nemotron 3 Ultra | 550B | 55B | Complex reasoning and long-running agentic systems | Up to 1M tokens |
B = billion parameters. “Active parameters” describes the approximate number used for a token's computation in a sparse Mixture-of-Experts model. It is not the same as the model's total storage size or a direct GPU-memory requirement. Nano's technical report specifies 3.2B active parameters, or 3.6B when embeddings are included. Super is described as a 120B / 12B model in its report, while Ultra is specified as 550B / 55B.
Nemotron 3 Nano
Choose Nano when the job needs capable reasoning but you want a smaller active compute footprint for specialised agent tasks, coding support, summarisation, information retrieval, or assistant workflows.
31.6B total · 3.2B active*Nemotron 3 Super
Consider Super for applications that coordinate multiple agents or handle sustained reasoning at higher volume, such as tool-driven workflows, coding tasks, and IT ticket automation.
~120B total · ~12B activeNemotron 3 Ultra
Consider Ultra when the application calls for the largest model in this core family, including complex research, planning, and long-running autonomous agent tasks.
550B total · 55B active*Nano's 3.6B figure includes embeddings; the technical report states 3.2B active per forward pass before embeddings are counted.
How does Nemotron 3 architecture work?
The core idea is a hybrid Mamba-Transformer Mixture-of-Experts (MoE) design. In plain language, NVIDIA combines different ways of processing information so the model can balance long-sequence efficiency with the attention mechanisms and specialised capacity needed for reasoning.
1. Mamba layers for sequence processing
Mamba-style state-space layers are designed to process sequences efficiently. They are part of the model's approach to handling long inputs and sustained workflows without relying only on standard attention layers.
2. Transformer attention for information relationships
Transformer attention helps a model relate different parts of the input. Nemotron 3 uses a hybrid arrangement rather than a purely Transformer-only backbone.
3. Sparse Mixture-of-Experts routing
MoE models contain specialised expert networks. A routing system selects a subset of experts for a given token rather than activating every expert for every token. This is why total parameters and active parameters are different figures.
Additional features in Super and Ultra
NVIDIA's Super and Ultra reports describe additional methods including LatentMoE, Multi-Token Prediction (MTP), and NVFP4 pretraining. MTP supports faster generation through speculative decoding. These features should not be assumed to be identical across every Nemotron checkpoint; check the relevant model report and model card.
What does the 1-million-token context window mean? It describes the maximum context length supported by the relevant model configuration. It does not mean every hosting provider, inference engine, or hardware setup will run the model at one million tokens by default. Runtime limits, memory use, latency, and cost depend on the serving configuration.
Which Nemotron 3 model should you choose?
Start with the workload rather than selecting the largest model automatically. A compact model can be a better fit for a focused sub-agent, while a larger model may be justified for more demanding reasoning tasks.
| Your project requirement | Model to evaluate first | Why |
|---|---|---|
| Low-cost, specialised agent tasks | Nemotron 3 Nano | Designed around inference efficiency for targeted workloads. |
| Summarisation, retrieval, coding support | Nemotron 3 Nano | Examples of Nano's stated target tasks include summarisation, information retrieval, and software debugging. |
| Several agents working together | Nemotron 3 Super | Positioned for collaborative agents and high-volume workflows. |
| Complex reasoning and long-running tasks | Nemotron 3 Ultra | The largest core model, designed for demanding reasoning and agentic applications. |
| Production deployment with a strict latency or cost target | Benchmark Nano, Super, and Ultra on your workload | Real performance depends on hardware, quantisation, concurrency, context length, and inference software. |
- Define the task.Write down whether your application needs retrieval, coding, tool use, planning, or multi-agent coordination.
- Set measurable targets.Choose acceptable latency, throughput, answer quality, context size, and serving cost before comparing models.
- Test the actual checkpoint.Use representative prompts and your own data. Compare quality and speed under the same hardware and inference settings.
- Check the deployment and licence details.Verify hardware requirements, supported inference frameworks, model terms, and data-handling requirements for the exact checkpoint.
Where can you download and run Nemotron 3?
Use NVIDIA's official research pages and model repositories as the starting points. The Hugging Face collection is useful for finding model checkpoints and related assets. Check the individual model card before downloading so you select the correct base, post-trained, or quantised version.
Official model and research links
- NVIDIA Nemotron v3 collection on Hugging Face — model checkpoints and related resources.
- Nemotron 3 family research page — family overview and links to the technical reports.
- Nemotron 3 Nano technical report (PDF).
- Nemotron 3 Super technical report (PDF).
- Nemotron 3 Ultra technical report (PDF).
Deployment options
NVIDIA's developer resources discuss deployment through open inference frameworks such as vLLM, SGLang, Ollama, and llama.cpp, as well as NVIDIA NIM services. Support can vary by model, checkpoint format, and framework version, so follow the current setup instructions for the exact model you intend to run.
Start with NVIDIA Developer: Nemotron AI models for technical resources, and NVIDIA's Nemotron platform page for the wider model ecosystem.
Is Nemotron 3 open source?
NVIDIA describes Nemotron 3 as a family of open models and releases model weights, datasets, and technical resources for the family. However, the phrase “open model” should not be treated as a substitute for reading the licence. Terms can differ by checkpoint, dataset, or associated asset.
Before using a model in a business or product, open the exact Hugging Face model card and licence file, review acceptable-use terms and redistribution conditions, and check whether any separately downloaded dataset or tool has its own licence. This page is an informational guide, not legal advice.
Important practical limits
- Parameter count is not a GPU estimate. Total weights, active parameters, precision, context length, KV cache, batch size, and parallelism all affect memory needs.
- Long context has a cost. A model's stated context limit does not guarantee that every provider can serve the full limit with acceptable latency or cost.
- Benchmark results need context. Compare the same task, prompt format, input/output lengths, hardware, quantisation, batch size, and inference engine.
- Model availability changes. Check the official model card and NVIDIA release pages for the latest checkpoint and deployment instructions.
Scope note: This guide focuses on the three core models in the original Nemotron 3 family—Nano, Super, and Ultra. NVIDIA's wider Nemotron ecosystem has since expanded to include additional specialised and 3.5-series models. Those should be checked separately rather than mixed into this three-model comparison. See the current NVIDIA Nemotron overview.
Frequently asked questions about NVIDIA Nemotron 3
What are the three NVIDIA Nemotron 3 models?
The core Nemotron 3 family consists of Nano, Super, and Ultra. Nano is positioned for efficient specialised tasks, Super for collaborative agents and high-volume workflows, and Ultra for more demanding reasoning and agentic applications.
What is the difference between Nemotron 3 Nano, Super, and Ultra?
The models differ in scale and intended workload. Nano has 31.6B total parameters and 3.2B active per forward pass (3.6B including embeddings). Super is approximately 120B total and 12B active. Ultra is 550B total and 55B active. Their best fit depends on the complexity and serving requirements of your application.
Does Nemotron 3 support a 1-million-token context window?
NVIDIA documents context lengths of up to one million tokens for the core Nemotron 3 models. Actual availability depends on the checkpoint, inference framework, provider configuration, and hardware.
Which Nemotron 3 model is best for AI agents?
There is no single best model for every agent. Nano is a good first model to evaluate for efficient specialised agents. Super is designed for collaborative agents and high-volume workflows. Ultra targets complex reasoning and long-running agentic tasks. Test them against your own workload before deciding.
Where can I download NVIDIA Nemotron 3?
Start with NVIDIA's Nemotron v3 collection on Hugging Face and follow the individual model cards for checkpoint downloads, files, and licence terms.
Can I run Nemotron 3 locally?
Local deployment depends on the model size, available GPU memory, quantisation, context length, and supported inference software. NVIDIA's developer resources list frameworks such as vLLM, SGLang, Ollama, and llama.cpp for Nemotron workflows, but check the current compatibility instructions for your exact checkpoint.
Is Nemotron 3 free for commercial use?
Do not assume every model, dataset, or checkpoint has identical terms. Read the licence attached to the exact asset and confirm that your intended commercial use is permitted before deployment.
Is Nemotron 3 the same as Nemotron 3.5?
No. This guide explains the core Nano, Super, and Ultra models in the Nemotron 3 family. NVIDIA's broader Nemotron portfolio has added later and specialised variants, including 3.5-series models. Check NVIDIA's current catalogue for those separate updates.
Sources and further reading
This guide prioritises NVIDIA's official research pages, technical reports, model collection, and developer documentation. The independent Medium article below is included as secondary background reading, not as the authority for model specifications.
- NVIDIA Research — NVIDIA Nemotron 3 Family of Models.
- NVIDIA Research — Nemotron 3 Ultra.
- NVIDIA — Nemotron foundation models.
- Hugging Face — NVIDIA Nemotron v3 collection.
- NVIDIA Developer — Nemotron AI models.
- NVIDIA Newsroom — Nemotron 3 family announcement.
- NVIDIA Technical Blog — Introducing Nemotron 3 Super.
- Independent Medium article — Engineering Efficiency for Agentic AI (December 19, 2025; supplementary reading).
Explore more on Nemotron3AI.com
Nemotron3AI.com is an independent informational website. It is not owned by, operated by, affiliated with, endorsed by, or sponsored by NVIDIA Corporation. NVIDIA, Nemotron, and related marks belong to their respective owners. Specifications and availability may change; always verify technical and licensing details with the official model documentation before deployment.