Nvidia unveiled a new suite of artificial intelligence (AI) models on Tuesday during its GPU Technology Conference (GTC) 2025. Named Llama Nemotron, these models focus on enhancing reasoning capabilities within large language models (LLMs), aimed at facilitating advanced AI workflows for developers and enterprises. The tech company, headquartered in Santa Clara, emphasized that these models can be utilized to create AI agents capable of functioning autonomously or collaboratively to tackle intricate tasks. Access to the Llama Nemotron models is currently provided through Nvidia’s platform and Hugging Face.
Nvidia Unveils New Reasoning-Focused AI Models
In a press release, Nvidia provided details on the new AI models. The Llama Nemotron reasoning models are built upon Meta’s Llama 3 series, incorporating enhancements from Nvidia’s post-training techniques. The company stated that these new models exhibit significant advancements in multistep mathematics, coding accuracy, reasoning capabilities, and complex decision-making processes.
Nvidia reported that the improvements achieved through this process have increased the models’ accuracy by as much as 20 percent in comparison to their base counterparts. Additionally, there has been a notable fivefold increase in inference speed relative to similar-sized open-source reasoning models. According to Nvidia, these capabilities enable the models to manage more sophisticated reasoning tasks, enhance decision-making efficiency, and lower operational costs for enterprises. This positions the LLM as a foundational tool for developing AI agents.
The Llama Nemotron reasoning models are offered in three parameter variations: Nano, Super, and Ultra. The Nano model is tailored for on-device and edge-based applications demanding high precision. The Super variant serves as a middle ground, providing robust accuracy and throughput on a single GPU. The Ultra model, designed for deployment on multi-GPU servers, boasts superior agentic accuracy.
The post-training of these reasoning models occurred on the Nvidia DGX Cloud, utilizing curated synthetic data generated from the Nemotron platform and other open models. Nvidia is also extending access to the tools, datasets, and post-training optimization techniques employed in the development of the Llama Nemotron models to the open-source community.
Nvidia is actively collaborating with enterprise partners to deliver these new models to developers and businesses. The reasoning models and NIM microservices can be utilized through Microsoft’s Azure AI Foundry, along with an option available via the Azure AI Agent Services. Additionally, SAP is integrating these models into its Business AI solutions and its AI copilot, Joule. Other companies leveraging Llama Nemotron models include ServiceNow, Accenture, and Deloitte.
The Llama Nemotron Nano and Super models, along with NIM microservices, are accessible to businesses and developers via an API through Nvidia’s platform, and also listed on Hugging Face here. They are provided under the permissive Nvidia Open Model License Agreement, allowing for both research and commercial applications.
