Photo credit: venturebeat.com
The field of agentic AI is advancing rapidly, with Nvidia unveiling a suite of services and models designed to enhance the creation and implementation of AI agents. This move positions Nvidia not just as a hardware leader but also as a key player in the burgeoning development of agentic AI technologies.
The new offering, dubbed Nemotron, consists of a variety of models derived from Meta’s Llama framework, utilizing Nvidia’s proprietary training methods and datasets. With Nemotron, Nvidia introduces three distinct model sizes: Nano, Super, and Ultra, each tailored for different applications. In addition to the Llama Nemotron optimized for language tasks, there is also a vision-oriented variant named Cosmos Nemotron, designed for physical AI applications. The parameters of these models vary significantly, with the Nano featuring 4 billion parameters, the Super 49 billion, and the Ultra boasting an impressive 253 billion parameters.
According to Nvidia, these models excel at various agentic tasks, including instruction following, chat interactions, function calling, coding, and mathematical computations. Rev Lebaredian, the Vice President of Omniverse and Simulation Technology at Nvidia, explained that each model size is optimized for specific computing environments. The Nano is tailored for cost-effective, low-latency applications suitable for PCs and edge devices, while the Super model is geared for high accuracy and throughput on single GPU setups. The Ultra model, on the other hand, is intended for data center environments where maximum accuracy is paramount.
“AI agents represent a digital workforce that not only performs tasks for us but collaborates alongside us,” Lebaredian remarked during a media briefing. With this in mind, the Nemotron model family is specifically designed for agentic AI applications.
These models are accessible via hosted APIs on platforms like Hugging Face and Nvidia’s own site, alongside availability through Nvidia’s AI Enterprise software platform for corporations.
Nvidia has a history of engaging with foundational models, having previously introduced a version of Nemotron called Llama-3.1-Nemotron-70B-Instruct, which demonstrated superior performance compared to similar offerings from OpenAI and Anthropic. The launch of NVLM 1.0, a family of multimodal language models, further underscores Nvidia’s commitment to advancing AI capabilities.
The growing trend of AI agents gained substantial traction in 2024 as businesses sought ways to integrate these systems into their operations. Experts believe this trend will persist in the current year, with prominent companies like Salesforce, ServiceNow, AWS, and Microsoft recognizing AI agents as the next pivotal evolution in generative AI within enterprises. AWS has recently incorporated multi-agent orchestration into its Bedrock platform, while Salesforce debuted Agentforce 2.0, expanding its agent accessibility to users.
However, for agentic systems to operate effectively, a robust orchestration framework is essential. Orchestration entails managing the interactions of multiple agents working across diverse systems.
Nvidia is now stepping into the realm of AI orchestration with its newly developed blueprints, which provide structured guidance for agents performing specific tasks. The company has formed partnerships with various orchestration platforms, including LangChain, LlamaIndex, CrewAI, Daily, and Weights and Biases, to create tailored blueprints on the Nvidia AI Enterprise framework. Each collaboration has led to the development of unique blueprints. For instance, CrewAI’s blueprint focuses on code documentation, ensuring easy navigation within code repositories. Similarly, LangChain has integrated Nvidia’s NIM microservices into its blueprint for structured report generation, facilitating varied output formats for internet search results.
“Successful orchestration of multiple agents is crucial for effectively deploying agentic AI,” Lebaredian stated. “These leading companies in AI orchestration are integrating all of Nvidia’s agentic components, including NIM, Nemo, and Blueprints, into their open-source platforms.”
Among the innovative blueprints, one designed to convert PDF content into podcasts aims to rival Google’s NotebookLM, while another blueprint focuses on creating agents capable of searching for and summarizing video content. Lebaredian emphasized that the intention behind the Blueprints initiative is to enable developers to swiftly deploy AI agents. To facilitate this, Nvidia has introduced Nvidia Launchables, a platform that allows developers to test, prototype, and run blueprints with a single click.
Looking ahead, orchestration may emerge as a significant development in 2025 as enterprises navigate the complexities of multi-agent environments.
Source
venturebeat.com