Run Nano through the official model card.

1. Review the model card and license

Read the intended use, supported languages, architecture notes, ethical considerations and NVIDIA Nemotron Open Model License before downloading weights.

2. Pick a serving route

Transformers is useful for direct Python evaluation. The official example uses the NVIDIA repository identifier and enables custom model code.

vLLM or SGLang can expose an OpenAI-compatible local endpoint for applications. The model card includes current launch examples for both.

Docker Model Runner offers a container-oriented route, while the model card also links to community quantizations for compatible local applications.

3. Validate your environment

Weights are large. Check storage, GPU memory, compute capability, framework versions and quantization options before downloading. Start with a short test prompt and measured output limit.

4. Keep reasoning configurable

Nano supports reasoning and non-reasoning behavior through its chat template. Use the exact current template from the official repository rather than copying an old snippet.