What Is NVIDIA Nemotron?
Nemotron is NVIDIA’s family of open models. What distinguishes it from most open-weight releases is the scope of what gets published: not only the weights, but the training data and the recipes and technical reports describing how the models were built. That matters for teams that need to audit what a model was trained on, or reproduce and adapt the training process rather than only fine-tune the result.
The models target agentic applications: tool use, multi-step reasoning, coding, document intelligence, and voice agents.
The Nemotron 3 Family
| Model | Size | Positioning |
|---|---|---|
| Nano | 30B | Cost-efficient sub-agents, with multimodal variants |
| Super | 120B | Efficiency and accuracy for multi-agent applications |
| Ultra | 550B | Frontier reasoning for complex enterprise workflows |
Specialized models sit alongside the main tiers, including Nemotron 3.5 Lightning (a 30B MoE) and Nemotron 3 Nano Omni (30B multimodal).
Architecture
Nemotron 3 uses a hybrid Mamba-Transformer mixture-of-experts architecture with context windows up to 1M tokens. The hybrid design is a throughput decision: state-space layers scale better with sequence length than full attention, and the MoE structure keeps active parameters well below total parameters, so serving cost tracks the active count rather than the headline size. For agentic workloads that push long contexts through many turns, both properties show up directly in serving economics.
Where to Get Them
Weights are published on Hugging Face and the models are available through build.nvidia.com, OpenRouter, a range of managed inference providers, and local tools such as Ollama, LM Studio, and llama.cpp.
Serving Nemotron
Because the weights are open, Nemotron models can be served on your own infrastructure with the same stack used for any other open model: vLLM, SGLang, or TensorRT-LLM as the engine, and NVIDIA Dynamo coordinating them across nodes when a model is too large for one node or needs disaggregated prefill and decode. MoE models in particular benefit from wide expert parallelism, which needs a high-bandwidth interconnect to be worth using.
For operators running a token factory, open weights are what make per-token pricing possible on your own terms: the model runs inside your boundary, on your hardware, metered and billed by you rather than resold from someone else’s API.
