Nvidia has unveiled its new family of open large language models (LLMs), known as the Llama Nemotron, which was announced on Monday. The company emphasized that the evolution of artificial intelligence (AI) agents necessitates the development of advanced AI models capable of managing agentic workflows. In a bid to enhance power and efficiency, Nvidia stated that these models are designed to create and deploy AI agents for a variety of applications, with access for enterprises through the Nvidia NIM microservice.
Nvidia Introduces Nemotron Family of AI Models
The announcement was made in a blog post, introducing the Nemotron series of open-source LLMs, which also includes the Cosmos Nemotron vision language models (VLMs). These VLMs are intended for the creation of AI agents capable of analyzing and responding to visual inputs such as images and videos. According to Nvidia, these vision-oriented agents can be utilized in various settings, including autonomous machinery, healthcare facilities, retail environments, and sports arenas.
The Nvidia Llama Nemotron models are constructed using Meta’s Llama foundation models and are optimized to facilitate the development of AI agents. Although the company did not disclose specific architectural details, it noted that the models are trained with “latest techniques and high-quality datasets.” They are geared towards enhancing agentic capabilities, encompassing tasks like instruction following, conversational engagement, function invocation, coding, mathematics, and beyond. The Nemotron family is designed to minimize the size of the AI agents, thus simplifying deployment.
Nvidia revealed that early adopters of the Llama Nemotron models will include SAP, ServiceNow, and various other platforms that focus on AI agents.
The models in the Nemotron and Cosmos Nemotron series will be available in three distinct parameter sizes: Nano, Super, and Ultra. The Nano model is tailored for cost-effectiveness and optimized for low-latency performance. The Super model offers high accuracy and can operate on a single GPU, while the Ultra model is engineered for high-accuracy demands suitable for data center applications.
Businesses will have the opportunity to access the Nemotron model family in two formats: as downloadable models and through the NIM microservice. Additionally, these models will be accessible via application programming interfaces (APIs). While the models are open-source, availability is restricted to academic and research purposes.
