Ways to Get Started With NVIDIA NeMo

NVIDIA NeMo is an agent-first, open suite of libraries and skills for accelerating AI agent specialization, optimization, and governance.

NeMo integrates with existing AI tools and agent frameworks to optimize specialized agents across any cloud, on-premises, or hybrid environment.

NeMo is available open source and supported as part of NVIDIA AI Enterprise. Pricing and licensing details can be found here.

NeMo can be used to customize large language models (LLMs), vision language models (VLMs),  automatic speech recognition (ASR), and text-to-speech (TTS) models.

NVIDIA AI Enterprise includes NVIDIA Business-Standard Support. For additional available support and services, such as NVIDIA Business-Critical Support, a technical account manager, training, and professional services, see the NVIDIA Enterprise Support and Service Guide.

NeMo Switchyard is an open source library that provides model-selection logic for AI agent workflows. It uses configured routing strategies to select from available models, helping teams balance quality, cost, and latency. Switchyard supplies the routing decision; the gateway or application sends the request to the selected model.

NeMo AutoModel is an open source, PyTorch-native training library for customizing day-zero Hugging Face models, including language and vision-language models. It provides configurable training recipes and distributed training capabilities to scale from experimentation to multi-GPU and multi-node workloads.

NeMo Megatron-Bridge is an open source library for pretraining and fine-tuning large models at scale. It combines Megatron Core’s distributed training capabilities with bidirectional Hugging Face checkpoint conversion, helping developers bring supported models into high-performance training workflows and export trained checkpoints back to Hugging Face format.

NeMo Curator is an open-source library that improves generative AI model accuracy by curating high-quality multimodal datasets. It consists of a set of Python modules expressed as APIs that make use of Dask, cuDF, cuGraph, and Pytorch to scale data curation tasks, such as data download, text extraction, cleaning, filtering, exact/fuzzy deduplication, and text classification to thousands of compute cores

NeMo Data Designer is a purpose-built microservice (also available as an open library) for AI developers that provides a programmatic way to generate synthetic data through configurable schemas and AI-powered generation models. It’s designed to integrate seamlessly into your AI development workflow.

NeMo Anonymizer is an open library that supports context-aware data anonymization within free text. After detecting sensitive entities, you can label, redact, hash, replace, or fully rewrite them.

NeMo Safe Synthesizer is an open library for generating synthetic versions of sensitive datasets. It creates entirely novel records with no one-to-one mapping to the original. It unlocks insights of your data while maintaining data privacy.

NeMo User Sim is an open source library that simulates diverse users interacting with models and agents through multi-turn conversations. It generates trajectories for evaluation, failure analysis, and targeted training, even before production traffic is available.

NeMo RL is an open source library for advanced reinforcement learning algorithms and scalable post-training to optimize and align AI agents at enterprise scale.

NeMo Gym is an open source library for building reinforcement learning (RL) training environments for large language models (LLMs). NeMo Gym provides infrastructure to develop environments, scale rollout collection, and integrate seamlessly with your preferred training framework. It helps generate data needed for RL training to equip AI agents and models with domain-specific skills.

NeMo User Sim generates diverse, multi-turn user interactions to uncover failures and produce evaluation and training data, even before production traffic is available. NeMo Gym provides verifiable environments in which models and agents perform tasks, enabling scalable evaluation and rollout generation for post-training and evaluation on agentic tasks. Use NeMo User Sim to explore user behavior and NeMo Gym to test and train agents in task environments.

NeMo Relay is an open source library that runs alongside your existing agent harness and captures a complete, structured record of everything your agent does—every tool call, LLM request, sub-agent interaction, and skill usage.

NeMo Compass is an open source analysis library that turns stored agent traces and evaluation evidence into a prioritized improvement backlog. It identifies recurring failures, links findings to supporting evidence, and highlights evaluation gaps so teams can decide what to fix and test next.

NeMo Eval Author is an open source library for creating domain-specific evaluations for agent workflows. It reviews an agent’s purpose, code, traces, and existing evaluations to identify coverage gaps, then creates custom tasks, metrics, and graders grounded in the work the agent performs.

garak is an open source vulnerability scanner for large language models. It uses automated probes to test for weaknesses such as prompt injection, sensitive information disclosure, and harmful outputs, helping teams identify safety and security issues and assess mitigations.

NeMo Fabric is an open source library that provides a consistent interface to configure, run, and collect results from agents across supported harnesses. It normalizes configuration, execution, and outputs so teams can integrate different harnesses without building a separate integration for each one.

NeMo Helix is an open source library for improving and hardening existing AI agents through repeatable evaluation, experimentation, optimization, and safety checks. It connects selected NeMo and third-party capabilities through modular plugins, helping teams test changes and compare quality, cost, latency, and safety before adopting improvements.

Use individual NeMo libraries when you need a specific capability in an existing workflow, such as data preparation, model training, evaluation, or guardrailing. Use NeMo Helix when you want to connect evaluation, experimentation, optimization, and safety checks into a repeatable agent improvement workflow. NeMo Helix brings selected NeMo and third-party capabilities together through modular plugins, so you can start with the tools you need and expand over time.

NVIDIA NIM, part of NVIDIA AI Enterprise, is an easy-to-use runtime designed to accelerate the deployment of generative AI across enterprises. This versatile microservice supports a broad spectrum of AI models—from open-source community models to NVIDIA AI Foundation models, as well as bespoke custom AI models. Built on the robust foundations of the inference engines, it’s engineered to facilitate seamless AI inferencing at scale, ensuring that AI applications can be deployed across the cloud, data center, and workstation.

Retrieval-augmented generation is a technique that lets LLMs create responses from the latest information by connecting them to the company’s knowledge base. NeMo works with various third-party and community tools, including Milvus, Llama Index, and LangChain, to extract relevant snippets of information from the vector database and feed them to the LLM to generate responses in natural language. Explore the AI Chatbot Using RAG Workflow page to get started building production-quality AI chatbots that can accurately answer questions about your enterprise data.

NVIDIA Blueprints are comprehensive reference workflows built with NVIDIA Nemotron open models and libraries. Each blueprint includes reference code, deployment tools, customization guides, and a reference architecture, accelerating the deployment of AI solutions like AI agents and digital twins, from prototype to production.

NVIDIA AI Enterprise brings together microservices, frameworks, and libraries for AI development with advanced GPU orchestration and infrastructure management in a fully supported, production-ready, commercial software suite. Deploy leading open source tools and AI models with confidence, improve productivity and time to value, and run AI workloads at scale with optimized resource utilization.