NVIDIA Nemotron Guide

Nemotron is NVIDIA’s family of open AI models. NVIDIA releases the model weights, and for Nemotron 3 also released trillions of tokens of training data, so developers can run, study, and customise the models themselves.

Nemotron is built for agents you control: open weights, open data, your hardware.

Nano

Small and fast, for local and edge use

Super

Mid-size, for most agent workloads

Ultra

The largest, for the hardest reasoning tasks

Nemotron 3 launched from December 2025. Third-party listings give different parameter counts, so check each model card on Hugging Face for exact sizes.

On August 11, 2026, NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model with about 3 billion active parameters, built for long-running local AI agents. It is on Hugging Face, including a compact NVFP4 version.

Reports of other Nemotron 3.5 models, such as a Super VL vision model, were still unclear at the time of writing.

Embedding

Nemotron 3 Embed in 1B and 8B sizes, for search and RAG

Speech

Streaming speech recognition models

Safety

Content safety models to filter inputs and outputs

• Hybrid mixture-of-experts: only part of the model runs for each token, so it is faster and cheaper
• Built for agent tasks such as tool use and long multi-step work
• Optimised for NVIDIA GPUs, including low-precision formats

• Hugging Face, under the nvidia organisation
• build.nvidia.com to try models in the browser and through an API
• NVIDIA NIM containers for production deployment
• Cloud and inference providers that host Nemotron

• Smaller models like Lightning can run on a strong workstation GPU
• Use quantised versions to save memory
• Tools such as llama.cpp, Ollama, and vLLM support many Nemotron models
• Check the model card for hardware needs

• Use a small model for routing and simple tool calls
• Use a larger model for planning and hard reasoning
• Pair it with an embedding model for retrieval
• Add a safety model in front of user-facing agents

Recent Nemotron models use NVIDIA’s open model licences, such as OpenMDW-1.1 for Ultra and the Nemotron 3.5 models. Read the licence on each model card before using it commercially.

Laptop or edge device

Nano or a quantised small model

Single workstation GPU

Lightning

Servers or cloud

Super or Ultra