NVIDIA launches the Llama Nemotron open reasoning models

Background

Open models are published so that others can download and adapt them, and reasoning models work through a problem step by step before giving an answer. NVIDIA, which is best known for the chips that train and run AI systems, had been building its own family of models on top of Meta's Llama 3 models, and in March 2025 it presented the result as a product line for developers and businesses.^[2]

What happened

On 18 March 2025, at its GTC conference, NVIDIA launched the Llama Nemotron family of open reasoning models. The family came in three sizes, Nano, Super and Ultra, and NVIDIA offered them as NIM microservices, which are packaged versions of models designed to be deployed on a company's own infrastructure.^[1] The announcement was framed around agentic AI platforms, meaning systems that carry out multi-step tasks with tools.^[1]

Key details

NVIDIA said that post-training lifted accuracy by up to 20% and that inference ran up to 5 times faster than other open reasoning models. Both figures came from NVIDIA's own announcement.^[1]

The launch named several large partners. Microsoft added the models to Azure AI Foundry, its platform for building AI applications, and SAP used the models in its Joule copilot, an AI assistant built into SAP business software.^[1] NVIDIA also said it would release the tools, datasets and post-training techniques openly, so that others could reproduce or extend the work.^[1]

Users could switch reasoning on or off at inference time, which is the moment a model is run rather than trained. This made it possible to use one model for both quick replies and slower, more careful reasoning.^[2] A technical report describing the work was also posted on arXiv.^[2]

Impact

The launch placed NVIDIA's open reasoning models in front of enterprise customers through cloud platforms and business software, not only in research settings. For developers, the release offered a set of openly available models of different sizes, with the tools and datasets NVIDIA promised to publish, to build reasoning features into their own products.^[1][2] Later NVIDIA model releases in the Nemotron line followed this pattern of publishing weights and training data together.