NVIDIA's open recipe book for training frontier models

What it does This is NVIDIA’s developer-facing companion to the Nemotron model family: a repository of training recipes, deployment cookbooks, and end-to-end examples for building with their open (weights, data, techniques) agentic AI models. It covers the full lifecycle—data curation, synthetic generation, pretraining, SFT, RL, evaluation, and optimization—through composable “steps” invoked via a nemotron steps CLI.

The interesting bit The “Nemotron steps” abstraction treats each training stage as a named, reusable unit with declared inputs/outputs and configurations. Recipes are just DAGs of these steps. There’s even a Claude Code plugin (nemotron-customize) that plans the DAG, validates artifact wiring, and emits YAML configs—turning a complex pipeline into a conversational build process. The catch: it only works when invoked from the exact repo root, which feels like a very 2025 kind of rough edge.

Key highlights

• Complete training recipes for Nemotron 3 Super (120B/12B active), Nano (31.6B/3.6B active), and Nano Omni (multimodal: text/image/video/audio)
• Steps catalog spans data curation, SDG, translation, pretraining, PEFT, RL (GRPO, RLHF, SWE-RL), checkpoint conversion, and benchmark evaluation
• Recipes use only open-source data subsets; NVIDIA explicitly notes results will differ from tech report benchmarks that used proprietary data
• Deployment support across NeMo, TensorRT-LLM, vLLM, SGLang, and NIM microservices
• Model tiers: Nano (edge/PC), Super (single GPU), Ultra (datacenter multi-GPU)

Caveats

• The open-source-data-only disclaimer means these are reference implementations, not reproductions of NVIDIA’s best results
• The Claude Code plugin’s path resolution is brittle—subdirectory invocation causes file read failures
• Heavy NVIDIA ecosystem dependencies: Megatron-Bridge, Automodel, NeMo-RL, NeMo Curator, DataDesigner, Evaluator

Verdict Worth bookmarking if you’re training or fine-tuning production LLMs and want to see how NVIDIA stages validation gates, curriculum learning, and multi-environment RL. Skip if you’re looking for a lightweight, framework-agnostic toolkit—this is thoroughly enmeshed with NVIDIA’s stack.