Mamba-2 layers
Mamba-style state-space layers are designed to process sequence information efficiently. In a hybrid model, they help handle long streams of tokens without using attention in every layer.
Looking for a clear guide to NVIDIA Nemotron 3 Nano? Learn how the 4B and 30B-A3B models differ, how the hybrid architecture works, what the 1-million-token context claim means, and how to download or run Nano with Ollama and other tools.
A straightforward explanation before you choose a checkpoint or deployment option.
Nemotron 3 Nano is NVIDIA’s smaller-model line in the Nemotron 3 family. It is designed for reasoning, instruction following, coding, tool use and agent-style workflows where efficient inference matters. The name now covers more than one size, so it is important to check which Nano checkpoint or runtime tag you are selecting.
The two sizes covered in current model resources are Nemotron 3 Nano 4B and the original Nemotron 3 Nano 30B-A3B. The 4B model is a separate compact model for lower-memory and edge-oriented deployments. The 30B-A3B model is a sparse Mixture-of-Experts (MoE) model with a much larger total parameter pool but only a smaller group of parameters used for each token.
Both belong to the Nano line, but they address different hardware and workload needs.
| Detail | Nemotron 3 Nano 4B | Nemotron 3 Nano 30B-A3B |
|---|---|---|
| Model size | 4 billion parameters | About 30B total; NVIDIA reports 31.6B in the technical report |
| Design goal | Compact local and edge AI workloads | More capable reasoning and agentic workloads with sparse activation |
| Architecture | Hybrid Mamba-Transformer MoE design | Hybrid Mamba-Transformer with sparse MoE |
| Where to start | NVIDIA’s 4B overview · Ollama 4B tag | Official BF16 checkpoint · Technical report |
| Ollama command | ollama run nemotron-3-nano:4b | ollama run nemotron-3-nano:30b |
Model availability, quantization, context limits and hardware needs vary by checkpoint and provider. Check the exact download or endpoint before deployment.
Numbers below refer to NVIDIA’s published technical report and model resources, not every third-party runtime’s default configuration.
| Developer | NVIDIA |
| Model type | Text language model with reasoning and non-reasoning modes |
| Total parameters | 31.6 billion |
| Active parameters | 3.2B per forward pass, or 3.6B when embeddings are included in the count |
| Training | Pretrained on 25 trillion text tokens, followed by supervised fine-tuning and large-scale reinforcement learning |
| Architecture | Hybrid Mamba-Transformer with sparse Mixture-of-Experts layers |
| Context | Up to 1 million tokens with a supported, configured serving setup |
| License | NVIDIA Nemotron Open Model License; review the exact terms before commercial use |
| Official checkpoint | NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 on Hugging Face ↗ |
The model combines different neural-network building blocks instead of relying on a standard Transformer stack alone.
Mamba-style state-space layers are designed to process sequence information efficiently. In a hybrid model, they help handle long streams of tokens without using attention in every layer.
Attention layers help the model relate different parts of the input. Nemotron 3 Nano mixes attention with Mamba-based processing rather than using only one approach.
An MoE layer has multiple specialist subnetworks, called experts. A router selects a small set for each token, so the model can have a large total parameter count without activating every expert at once.
NVIDIA reports supervised fine-tuning and large-scale reinforcement learning across diverse environments. These stages are intended to improve instruction following, reasoning and agentic behaviour.
NVIDIA’s paper describes a hybrid Mamba-Transformer MoE model. The exact layer design and active-parameter accounting should be taken from the technical report and the selected checkpoint’s model card rather than inferred from the short model name.
Yes, NVIDIA documents support up to 1M tokens for the 30B-A3B model, but the number you can use depends on the runtime and configuration.
A context window is the amount of text the model can consider in a single interaction, including the prompt, conversation history and generated tokens. A larger context can help with long documents, code repositories and multi-step workflows.
Do not assume that every website or API offering Nemotron 3 Nano exposes the full 1M-token limit. NVIDIA’s Hugging Face example uses 256K and shows how to enable a 1M maximum in vLLM with an additional environment variable and model-length setting. NVIDIA NIM and third-party gateways may list a lower endpoint limit, such as 262K.
Reference: Hugging Face model card and vLLM setup ↗ · NVIDIA NIM model page ↗.
Benchmarks can help you understand a model’s strengths, but scores should be read with the task, test setup and model configuration in mind.
In its published evaluation, NVIDIA reports results across general knowledge, reasoning, science and coding benchmarks. The selected results below are from the official Hugging Face model card for the post-trained 30B-A3B checkpoint.
| Benchmark | Nemotron 3 Nano 30B-A3B | Qwen3-30B-A3B-Thinking-2507 | GPT-OSS-20B |
|---|---|---|---|
| MMLU-Pro | 78.3 | 80.9 | 75.0 |
| AIME25 (no tools) | 89.1 | 85.0 | 91.7 |
| GPQA (no tools) | 73.0 | 73.4 | 71.5 |
| LiveCodeBench | 68.3 | 66.0 | 61.0 |
These are NVIDIA-published evaluation results, not independent tests by Nemotron3AI.com. Scores are not directly interchangeable across different prompts, tools, model versions or evaluation harnesses. The table also shows why it is not accurate to say Nano wins every benchmark.
Source: NVIDIA model card benchmark table ↗ · Technical report ↗
Potential workloads depend on the selected size, quality requirements and deployment environment.
Build assistants that can choose tools, follow structured steps and work through bounded tasks. Always validate tool permissions and outputs.
Explore code explanations, generation, refactoring and repository questions. Test generated code before using it in production.
Use supported long-context configurations to review large documents or extended conversations, while checking the actual endpoint context limit.
Run a suitable checkpoint on infrastructure you control. Local deployment may help with data control, but security and retention depend on your own setup.
Ollama currently lists both a 4B and a 30B Nano tag. Choose the one your machine can support.
Use this tag to pull and start the smaller 4B model:
ollama run nemotron-3-nano:4bUse this tag for the larger 30B-A3B model. Download size and memory needs depend on the selected quantized artifact and runtime.
ollama run nemotron-3-nano:30bThe commands above are copied from the current Ollama model instructions. Tags and available variants can change, so check Ollama’s model page if a command stops working.
Use the source that matches your goal: official model weights, local runtime, technical research or hosted inference.
A useful model guide should explain what to verify—not only list the benefits.
The 30B-A3B model activates only a fraction of its parameters for each token, but that does not mean it only needs memory for 3B parameters. The full checkpoint, runtime buffers, KV cache and chosen context length all affect the hardware needed. Quantized downloads can reduce weight storage, but quality and performance should be tested on your workload.
The 1M-token capability is a maximum described by NVIDIA for a configured setup. Hosted APIs and serving frameworks may enforce a smaller limit. Always check the provider’s current model page and API documentation.
The 30B-A3B Hugging Face model card points to the NVIDIA Nemotron Open Model License. Read the applicable license and any linked terms before redistributing weights, building a commercial product or using the model in a regulated workflow. “Open weights” should not be interpreted as public domain or unrestricted use.
Like other language models, Nano can produce incorrect or incomplete answers. Validate generated code, calculations, citations and tool actions. For consequential decisions, require appropriate human review and test the model with representative data before deployment.
Quick answers to common questions about model size, setup, context and licensing.
Nemotron 3 Nano is NVIDIA’s Nano model line in the Nemotron 3 family. Current resources include a compact 4B model and the original 30B-A3B sparse MoE model, designed for reasoning, coding, instruction following and agentic workloads.
No. They are distinct model sizes. The 4B model is a compact checkpoint, while the 30B-A3B model has about 31.6B total parameters and activates a smaller subset per token. A 4B model is not simply the 30B model with quantization applied.
It refers to a model with roughly 30 billion total parameters and around 3 billion parameters activated for each token by its sparse Mixture-of-Experts architecture. NVIDIA’s technical report gives the more precise count as 31.6B total and 3.2B active, or 3.6B including embeddings.
The 30B-A3B model can support up to 1M context tokens with the correct serving configuration. Some model cards, NVIDIA endpoints and third-party providers expose lower defaults or limits, so check the deployment you plan to use.
For the current Ollama library tags, use ollama run nemotron-3-nano:4b for the compact model or ollama run nemotron-3-nano:30b for the 30B model. Check the Ollama library page for current tags and requirements.
The model card specifies the NVIDIA Nemotron Open Model License. Whether a particular use is permitted depends on the license terms and your use case. Review the current license before commercial deployment or redistribution.
No. Nano Omni is a separate multimodal model. NVIDIA’s NIM listing describes it as accepting video, audio, image and text inputs. The original Nano 30B-A3B checkpoint covered in this guide is a text model.
Continue from this model-specific guide to the family overview and other topics on Nemotron3AI.com.
Open NVIDIA’s official Hugging Face model card for checkpoint details, setup notes, evaluation results and license information.
Technical facts were checked against NVIDIA material and current model-provider documentation on 2 November 2026.
Nemotron3AI.com is an independent informational website and is not affiliated with, endorsed by or sponsored by NVIDIA Corporation. NVIDIA and Nemotron are trademarks of NVIDIA Corporation. This guide summarises public documentation; it does not claim independent benchmark testing. Model availability, pricing, licenses and provider limits may change. Always confirm details with the relevant official source before deployment.