Heads up This site is currently under heavy development.
← all tools
◆ AI Model & Data Infrastructure

Nemotron

open-source

Summary

Nemotron is an open-source (Apache 2.0) collection of training recipes for NVIDIA's Nemotron family of agentic AI models, covering data preparation, pretraining, supervised fine-tuning, and reinforcement learning stages. It runs from the command line: you clone the repository and drive each pipeline stage with the `nemotron` CLI, submitting jobs to a Slurm cluster through NeMo-Run. It is aimed at ML engineers and researchers building or adapting open models rather than consuming them as a hosted API, and it sits in the open-models and training-infrastructure space alongside NVIDIA's own NeMo stack (Megatron-Bridge, NeMo-RL), which it's built on, with artifact lineage tracked through Weights & Biases. First committed in 2025, it already has 47 contributors and nearly 540 commits in the past year, indicating active development.

What Nemotron answers

Do I need my own Slurm cluster to run this, or can I try it on a single machine?

every pipeline stage submits jobs through NeMo-Run to a configured Slurm cluster, so you need that infrastructure in place before any stage runs

Which models can I actually train with this today?

recipes cover Nemotron 3 Nano, Nemotron 3.5 Lightning, Nemotron 3 Nano Omni, and Nemotron 3 Ultra, each with its own pretraining, SFT, and RL stages

Can I start from a pretrained checkpoint instead of training from scratch?

checkpoints for each model are published on Hugging Face, so you can skip pretraining and run only the SFT or RL stages

What do I get once training finishes — just weights, or something deployable?

the usage cookbook covers taking a trained model into TensorRT-LLM, vLLM, SGLang, NIM, or Hugging Face for serving

How do I know which data and config produced a given checkpoint?

each stage's inputs and outputs are recorded as tracked artifacts in Weights & Biases, linking a model back to the data blend and config that built it

Do I have to write my own training data pipeline first?

data preparation is a distinct stage in each recipe, driven by the same command-line tool as pretraining, SFT, and RL

Release history

Nothing shipped since we started watching. Releases and docs updates land here as they happen.

watching since Jul 25, 2026

my-toolchain — 0 tools
paste an install list to detect your tools

A brew list, a Brewfile, requirements.txt, a Dockerfile — or just the product names, free-form. Nothing leaves your browser.

    browse all tools →