NVIDIA launches family of open reasoning AI models for developers and enterprises to build agentic AI platforms

NVIDIA has announced the introduction of the Large Language Model Meta AI (LLaMA) Nemotron family of models, which feature advanced reasoning capabilities aimed at developers and enterprises seeking to create Artificial Intelligence (AI) agents. These models are designed to assist with complex tasks, either autonomously or collaboratively.

The new LLaMA Nemotron models, built on the existing LLaMA architecture, have undergone post-training enhancements aimed at improving their performance in areas such as multistep math and complex decision-making. The post-training process has reportedly boosted the models' accuracy by up to 20% compared to their base version and increased their inference speed by five times when compared to other available open reasoning models. This improved performance allows the models to address more complex reasoning tasks, thereby supporting better decision-making and potentially lowering operational costs for enterprises.

Companies including Accenture, Amdocs, Atlassian, Box, Cadence, CrowdStrike, Deloitte, IQVIA, Microsoft, Situational Awareness Platform (SAP), and ServiceNow are collaborating with NVIDIA to leverage these new reasoning models as they integrate them into their offerings. For instance, Microsoft plans to incorporate these models into its Azure AI Foundry, enhancing services like the Azure AI Agent Service for Microsoft 365. SAP is using them to advance its Business AI solutions and improve code completion accuracy in its programming environments.

According to NVIDIA's CEO Jensen Huang, the enhancements provided by their open reasoning models and tools equip enterprises with the necessary resources to develop a more capable AI workforce. The LLaMA Nemotron models are accessible through NVIDIA NIM microservices, available in various configurations to suit different deployment scenarios. Each size—Nano, Super, and Ultra—targets distinct operational requirements, offering options from high accuracy for Process Control System (PCS) to maximum capability on multi-GPU setups.

NVIDIA has also specified that the datasets and post-training optimization techniques used for these models will be available openly, facilitating custom model development by enterprises. The models can be deployed in production environments utilizing NVIDIA AI Enterprise, with future plans to release additional toolkits aimed at improving the functionality of AI agents.

The LLaMA Nemotron models are available for testing and development via a hosted Application Programming Interface (API) from NVIDIA, with free access for members of the NVIDIA Developer Program. Key components like the NVIDIA AI-Q Blueprint are set to be released in April, alongside the AgentIQ toolkit available now on GitHub.