NVIDIA Nemotron Developer Repository

Open and efficient models for agentic AI. Training recipes, deployment guides, and use-case examples for the Nemotron family.

🎉Nemotron 3 Ultra was announced at GTC San Jose 2026. The model is open-source on Hugging Face, and the training recipe is now available in this repo. To learn more, see the usage guide!

🎉Nemotron 3 Nano Omni is now released — a 30B-A3B hybrid Mamba-Transformer MoE with native text, image, video, and audio support, designed as a multimodal perception sub-agent for agentic AI. See the release blog, the training recipe, and the model weights.

Why Nemotron?

Use from Claude Code

This repo ships a Claude Code plugin called nemotron-customize that turns the step catalog under src/nemotron/steps/ into a guided, repo-native pipeline builder.

Install once:

/plugin marketplace add NVIDIA/Nemotron
/plugin install nemotron-customize@nvidia-nemotron

Then, start Claude Code from the repo root and invoke the skill:

cd /path/to/Nemotron        # repo root: must contain pyproject.toml and src/nemotron/steps/
claude

/nemotron-customize

The skill resolves all file paths against your current working directory, so it must be invoked from the Nemotron checkout root. Running it from a subdirectory will cause file reads to fail.

The skill plans the step DAG, validates artifact wiring, and emits the YAML configs needed to run the requested pipeline. See skills/nemotron-customize/SKILL.md for the full contract.

The marketplace installs only nemotron-customize. The other folders under skills/ (model knowledge bases, contributor add-* skills) stay on disk for repo browsing but are not loaded as plugins.

Repository Overview

nemotron/
│
├── src/nemotron/steps/      Modular building blocks for training, eval, SDG, and more
│
├── src/nemotron/recipes/    Training recipes (complete, reproducible pipelines)
│
├── usage-cookbook/          Usage cookbooks (deployment and model usage guides)
│
└── use-case-examples/       Examples of leveraging Nemotron in agentic workflows

Which section should I use?

What is Nemotron?

NVIDIA Nemotron is a family of open, high-efficiency multimodal models purpose-built for agentic AI.

Model Tiers:

• Nano — Optimized for edge and PC deployments
• Super — Single GPU deployment with highest throughput
• Ultra — Multi-GPU datacenter applications

Nemotron models excel at coding, math, scientific reasoning, tool calling, instruction following, and visual reasoning. Deploy across edge, single GPU, or data center environments with support for NeMo, TensorRT-LLM, vLLM, SGLang, and NIM microservices.

Nemotron Steps

A Nemotron step is a named, reusable unit of work that you invoke with the nemotron steps CLI. Each step packages a description of the work it performs, the artifacts it consumes and produces, and one or more named configurations that supply parameter values. Steps live under src/nemotron/steps/, and the CLI discovers them at startup.

The training recipes in the next section are composed from these steps. Run a step on its own when you want one stage, or chain steps together when you need a different pipeline shape than the published recipes.

Step Categories

The catalog covers the full training lifecycle.

• Data curation and preparation with curate/* and data_prep/*.
• Synthetic data generation (SDG) with sdg/*.
• Corpus translation with translate/*.
• Bring-your-own benchmark generation with byob/*.
• Pretraining, supervised fine-tuning (SFT), parameter-efficient fine-tuning (PEFT), and reinforcement learning (RL) with pretrain/*, sft/*, peft/*, and rl/*.
• Checkpoint conversion and model optimization with convert/* and optimize/*.
• Benchmark evaluation with eval/*.
• Execution-profile setup with env/*.

Documentation

• About Nemotron Steps is the entry point for the step model.
• Nemotron Steps Basics defines step, configuration, environment profile, and artifact.
• Getting Started With Steps walks through the CLI on a small example.

Training Recipes

The Nemotron repository provides reproducible training pipelines from raw data to deployment-ready models. These implementations reflect how large language models are actually trained: careful experimentation, validation gates, and systematic optimization.

Why Complete Pipelines?

Training a production model involves interconnected components. Isolated examples miss how stages interact. Complete pipelines show:

• How data quality affects downstream performance across pretraining, SFT, and RL
• Which training techniques actually work together, not just in theory
• Where validation gates prevent failures and maintain reproducibility
• How to balance competing objectives across stages

Because these are complete systems, you can extract specific techniques with confidence. Each component has been proven to work in context.

Each Recipe Includes

• 🎨 Synthetic Data Generation - Scripts to generate synthetic datasets using NVIDIA-NeMo/DataDesigner
• 🗂️ Data Curation - Scripts to prepare training data using NVIDIA NeMo Curator for scalable data processing, filtering, and quality enhancement
• 🔁 Training - Complete training loops with hyperparameters using:
  • NVIDIA-NeMo/Megatron-Bridge for Megatron models
  • NVIDIA-NeMo/Automodel for HuggingFace models
  • NVIDIA-NeMo/NeMo-RL when RL is needed
  • Includes GPU-accelerated last-mile data processing (tokenization + optional sequence packing) for optimal training efficiency
• 📊 Evaluation - Benchmark evaluation on standard suites using NVIDIA NeMo Evaluator
• 📖 Documentation - Detailed explanations of each stage

Available Recipes

Nemotron 3 Ultra

A training recipe for NVIDIA's largest Nemotron 3 model — a 550B-A55B hybrid Mamba-Attention Mixture-of-Experts Transformer with LatentMoE and multi-token prediction (MTP), pretrained in NVFP4 and extended to 1M-token context for datacenter-scale agentic reasoning.

Open-Source Data Only: These recipes train exclusively on the open-sourced subset of training data. Results will differ from the tech report benchmarks, which used additional proprietary data. Use these recipes as reference implementations to apply the methodology with your own data.

Model Specifications:

• 550B total / 55B active parameters (MoE)
• Hybrid Mamba-Attention architecture with LatentMoE + two shared-weight MTP layers
• 20T pretraining tokens in NVFP4, two-phase data curriculum
• Up to 1M (1,048,576) context length
• Full program: Pretrain → SFT → RLVR → MOPD → MTP Boosting (this recipe covers Pretrain → SFT)

What You Can Extract:

• Two-phase pretraining data mixture (tech-report Figure 4) over the open Nemotron datasets
• Ray-based data prep: tokenize raw datasets → Megatron bin/idx (pretrain) and pack chat data → Parquet (SFT)
• New open pretraining datasets: Specialized-v1.2 (Multiple-Choice / Generative / Fact-Seeking / Moral-Scenarios) and Legal-v1
• Stage-local container builds (Day-0 Megatron-Bridge) for both pretrain and SFT
• Megatron-Bridge training at Ultra scale (TP=2 / PP=12 / EP=32 pretrain, PP=6 SFT)

Resources:

• Training Guide
• Usage Guide
• Model Weights (BF16)

Nemotron 3 Super

A complete training recipe for the frontier Hybrid Mamba Latent Mixture-of-Experts Transformer model with state-of-the-art reasoning, coding, and agentic capabilities.

Open-Source Data Only: These recipes train exclusively on the open-sourced subset of training data. Results will differ from the tech report benchmarks, which used additional proprietary data. Use these recipes as reference implementations to apply the methodology with your own data.

Model Specifications:

• 120B total / 12B active parameters
• Multi-stage RL pipeline: 3× RLVR + 2× SWE-RL + RLHF across 21 reward environments
• Asynchronous GRPO with decoupled training and inference

What You Can Extract:

• Large-scale pretraining with data curriculum
• Multi-domain SFT pipeline
• Multi-environment RLVR with 21 simultaneous reward environments
• SWE-RL with container-isolated sandbox execution
• GenRM-based RLHF with principle-following rewards
• Asynchronous GRPO at 1K GPU scale

Resources:

• Training Guide
• Tech Report
• Model Weights (Instruct)

Nemotron 3 Nano

A complete training recipe for the open, efficient Mixture-of-Experts hybrid Mamba-Transformer model optimized for agentic reasoning.

Open-Source Data Only: These recipes train exclusively on the open-sourced subset of training data. Results will differ from the tech report benchmarks, which used additional proprietary data. Use these recipes as reference implementations to apply the methodology with your own data.

Model Specifications:

• 31.6B total parameters, 3.6B active per forward pass
• 25 trillion pretraining tokens with curriculum learning
• Up to 1M context length
• 3.3x higher inference throughput than similarly sized models

What You Can Extract:

• Curriculum-based pretraining with two-phase data mixture
• Long-context extension via CPT methodology
• Multi-domain SFT with 12+ data sources
• InfinityByte cross-domain code synthesis
• Tool-calling fine-tuning and budget-controlled reasoning
• Multi-environment RLVR with GRPO
• GenRM reward modeling with circular comparison
• DPO for tool hallucination reduction

Resources:

• Training Guide
• Tech Report
• Model Weights (Base)
• Model Weights (Instruct)
• Model Weights (FP8)

Nemotron 3 Nano Omni

A multimodal training recipe for the 30B-A3B hybrid Mamba-Transformer Mixture-of-Experts model. Native support for text, image, video, and audio in a single decoder, designed as a perception sub-agent for agentic AI.

![Nemotron 3 Nano Omni hybrid MoE architecture: each modality (audio via Parakeet, vision via C-RADIOv4-H + 3D convolution + Efficient Video Sampling