NVIDIA debuts the Nemotron 3 family of open models

Background

Nemotron 3 is a family of open models from NVIDIA's research group. The announcement made the open weights available together with the training data and methods behind them.^[2]

What happened

On 15 December 2025 NVIDIA announced the Nemotron 3 family of open models. The family came in three sizes, Nano, Super and Ultra. Nano was available immediately, while Super and Ultra were expected in the first half of 2026.^[1] NVIDIA described the sizes in approximate terms at the time: Super at about 100 billion parameters and Ultra at about 500 billion.^[1]

Key details

The models use a hybrid latent mixture-of-experts architecture. A mixture-of-experts model activates only some of its expert sub-networks for each token, which keeps the cost of each answer lower than the total size of the model would suggest. Super and Ultra are trained in NVIDIA's 4-bit NVFP4 format on Blackwell, the company's current GPU architecture.^[1]

Nano has a context window of one million tokens and, according to NVIDIA, up to four times the token throughput of Nemotron 2 Nano. The context window is the amount of text the model can consider at one time.^[1]

NVIDIA released three trillion tokens of new training datasets, an agentic safety dataset and the open-source NeMo Gym library for reinforcement learning.^[1] The agentic safety dataset is aimed at systems that take actions with tools, and NeMo Gym provides environments in which models can be trained with reinforcement learning. The weights, training recipe and data were published, along with FP8 and BF16 checkpoints.^[2]

Impact

The announcement made a complete open package available. Developers could download the Nano model right away and study its recipe, while the larger Super and Ultra models were scheduled for later in the year. Publishing the training data alongside the weights meant that outside groups could inspect how the models were built and reproduce parts of the process.^[1][2]