Developer companion repo for working with NVIDIA's Nemotron models: inference, fine-tuning, agents, visual reasoning, deployment.

nemotron/
│
├── usage-cookbook/        Usage cookbooks (how to deploy, and simple model usage guides)
│
│
└── use-case-examples/     Examples of leveraging Nemotron Models in Agentic Workflows and more 

NVIDIA Nemotron™ is a family of open, high-efficiency models with fully transparent training data, weights, and recipes.

Nemotron models are designed for agentic AI workflows — they excel at coding, math, scientific reasoning, tool calling, instruction following, and visual reasoning (for the VL models).

They are optimized for deployment across a spectrum of compute tiers (edge, single GPU, data center) and support frameworks like NeMo and TensorRT-LLM, vLLM, and SGLang, with NIM microservice options for scalable serving.

• Usage Cookbook - Practical deployment and simple model usage guides for Nemotron models
• Use Case Examples - Practical use-case examples and apps
• Nemotron Developer Page - Developer resources for the Nemotron family of models
• Nemotron Research Hub - Research affiliated with the Nemotron effort
• Nemotron Datasets - Datasets part of various Nemotron collections, from pre-training to post-training

Have an idea for improving Nemotron models? Visit the Nemotron Ideas Portal to:

• 🗳️ Vote on existing feature requests
• 💭 Submit your own ideas and suggestions
• 📊 See what the community is requesting

Your feedback helps shape the future of Nemotron models!

Full, reproducible training pipelines will be included in the nemotron package at src/nemotron/recipes/.

• 🎨 Synthetic Data Generation - Scripts to generate synthetic datasets using NVIDIA-NeMo/DataDesigner
• 🗂️ Data Curation - Scripts to prepare training data using NVIDIA-NeMo/Curator
• 🔁 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
• 📊 Evaluation - Benchmark evaluation on standard suites using NVIDIA-NeMo/Evaluator
• 📖 Documentation - Detailed explanations of each stage

Learn how to deploy and use the models through an API.

Below is an outline of the end-to-end use case examples provided in the use-case-examples directory. These scenarios demonstrate practical applications that go beyond basic model inference.

• Agentic Workflows
  Orchestration of multi-step AI agents, integrating planning, context management, and external tools/APIs.

• Retrieval-Augmented Generation (RAG) Systems
  Building pipelines that combine retrieval components (vector databases, search APIs) with Nemotron models for grounded, accurate outputs.

• Integration with External Tools & APIs
  Examples of Nemotron models powering applications with structured tool calling, function execution, or data enrichment.

• Production-Ready Application Patterns
  Architectures supporting scalability, monitoring, data pipelines, and real-world deployment considerations.

See the use-case-examples/ subfolders for in-depth, runnable examples illustrating these concepts.

We welcome contributions! Whether it's examples, recipes, or other tools you'd find useful.

Please read our Contributing Guidelines before submitting pull requests.

• Contributing Guidelines - How to contribute to this project
• Changelog - Version history and changes

Apache 2.0 License - see LICENSE file for details.

NVIDIA Nemotron - Open, transparent, and reproducible.