Llama Nemotron (Nano, Super, Ultra)

What it is

Llama Nemotron is a family of open reasoning models from NVIDIA. NVIDIA launched it on 18 March 2025 at its GTC conference, offering three sizes: Nano, Super and Ultra.^[1] “Open reasoning” means two things here: the models are built to work through problems step by step before answering, and their weights are published for others to use.^[2] The family's sizes are Nano at 8 billion parameters, Super at 49 billion and Ultra at 253 billion. All three are derived from Meta's Llama 3 models.^[2]

Origins

The models were built in several stages. Neural architecture search was used to find designs that run faster at inference time, meaning when the model is answering a question. Knowledge distillation, in which a smaller model learns to imitate a larger one, was applied next. Continued pre-training followed, then supervised fine-tuning, and finally large-scale reinforcement learning aimed at improving reasoning.^[2]

Main features

The most distinctive feature is a dynamic reasoning toggle. The same model can run in a standard chat mode or in a reasoning mode, and the choice is made at inference time, when the model is used rather than when it is trained.^[2]

NVIDIA gave specific performance claims. It said post-training improved accuracy by up to 20% over the base models, and that inference ran up to 5 times faster than other leading open reasoning models.^[1] These figures are NVIDIA's own claims and were not independently verified in the announcement.

Each size was positioned for a different setting. Nano was aimed at PCs and edge devices, which are computers or devices located close to where data is produced. Super was described as giving the best accuracy and throughput on a single GPU. Ultra was positioned for maximum agentic accuracy on multi-GPU servers, where “agentic” refers to systems that carry out multi-step tasks with tools.^[1]

The models were released under the NVIDIA Open Model License, and the release included a dataset called Llama-Nemotron-Post-Training-Dataset.^[2] Nano and Super were available at launch on build.nvidia.com and on Hugging Face, a public platform for sharing models. The Ultra model followed in April 2025 as Llama-3.1-Nemotron-Ultra-253B-v1.^[1][3]

Where it fits

Llama Nemotron reached businesses through a set of partners. NVIDIA named Accenture, Amdocs, Atlassian, Box, Cadence, CrowdStrike, Deloitte, IQVIA, Microsoft (through Azure AI Foundry), SAP (through its Joule assistant) and ServiceNow.^[1] The partners ranged from software companies such as Atlassian and ServiceNow to consulting firms such as Deloitte and IQVIA, a company that works with health and life-sciences data. The partner list shows the family was aimed at enterprise software and services as well as at individual developers, and that businesses were expected to build their own products and services on top of the models.

The technical report, which describes the training methods in detail, was posted on 2 May 2025.^[2]