NVIDIA: the GPU giant and its open Nemotron models

From video games to AI

NVIDIA was founded in 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem, with the goal of bringing 3D graphics to gaming and multimedia. Based in Santa Clara, California, it has become the leading supplier of the GPUs used to train and run AI models. It is firmly a big-tech company, yet its work as a model maker is far less known to the general public.

Nemotron: a family built for agents

Nemotron models are large language models offered in three tiers: Nano for fast tasks and specialized sub-agents, Super for a balance of accuracy and throughput, and Ultra for the most demanding reasoning. Recent generations combine two architectures (Transformer and Mamba) with a mixture of experts, an approach aimed at generation speed. As NVIDIA puts it, “faster AI is smarter AI.”

The lineup also includes more specialized models: multimodal models, retrieval models (embedding and reranking), language-specific versions, and the Llama Nemotron line, built on top of Meta’s Llama models.

Openness that goes beyond weights

Many labs release open weights without explaining how the model was made. NVIDIA goes further by also publishing its training datasets and techniques, and that is what sets it apart most. For researchers or businesses that care about traceability, this makes it possible to understand what is in the model, retrain it or adapt it.

The license, the NVIDIA Open Model License, is described as permissive: commercial use, modification and redistribution are allowed, with no requirement to credit NVIDIA. It is not a standard open-source license, though, so read it before any sensitive project.

How to access Nemotron

• Hugging Face, to download the weights and run them locally.
• build.nvidia.com, to try the models in your browser and prototype through the API.
• NVIDIA NIM, production-ready microservices under an NVIDIA AI Enterprise license.

To compare Nemotron with other open models, visit our open-source models page.