Nemotron

A family of AI models for self-hosted agents, with open weights and training data, supporting Ollama and llama.cpp on NVIDIA GPUs.

Nemotron is NVIDIA's family of AI models for developers building agents that reason, write code and use tools. Teams can evaluate and adapt the models using published weights, training data and training recipes.

Reasoning and tool use: Models support coding, math and multi-step planning, with Lightning focused on frequent specialized tasks in long-running agents.
Multimodal understanding: Nano Omni handles text, images, video and audio in one model for applications such as document understanding and computer use.
Document retrieval: Retriever models extract information, create embeddings and rerank results; Parse extracts text and tables from complex layouts, including multi-column documents.
Speech and safety: Speech models support transcription, speech generation and translation, while Safety models handle multilingual moderation, jailbreak detection and personal information detection.
Model customization: Training recipes cover pretraining, fine-tuning and reinforcement learning, with data preparation and evaluation tools for adapting models to specific tasks.

Self-hosted deployment supports Ollama, llama.cpp, vLLM, SGLang and TensorRT-LLM on NVIDIA GPUs. Nano targets PCs and edge devices, Super supports a single data center GPU, and Ultra targets multi-GPU systems. NVIDIA NIM provides inference microservices for deployment on GPU-accelerated systems. OpenRouter and managed inference providers offer hosted access, where inference runs outside your hardware.

The developer repository uses the Apache 2.0 license.