NVIDIA is fully committed to developing AI Agent! Open source exclusive large model, the strongest AI factory

Early this morning, Nvidia, the global leader in AI, held the "GTC 2025" global product and developer conference in San Jose, USA, covering technology sharing such as large models, AI Agents, physical robots, AI-specific chips, and quantum computing.

Among them, NVIDIA announced on its official website that it has reached a technical cooperation with Oracle, providing more than 160 native AI tools and more than 100 microservices to help companies build ultra-complex, seamless Agent applications and automate complex and repetitive digital businesses.

A basic model dedicated to Agent, Llama Nemotron , has also been open-sourced. Currently, Microsoft, Accenture, Deloitte, SAP, and ServiceNow are working with NVIDIA to develop commercial agents based on this model.

At the same time, Blackwell Ultra, an AI factory dedicated to training and reasoning of AI Agents, was released, with performance significantly improved over the previous generation.

The GTC 2025 conference will be held for 5 days, and it is expected that there will be more than 1,000 product and technology sharing sessions. However, NVIDIA put AI Agent as the highlight on the first day, and introduced it in a series of press releases, just like yesterday's Zoom, which is enough to show the importance of intelligent agents.

Now, the "AIGC Open Community" will introduce to you what Agent feasts NVIDIA has prepared for the opening ceremony on the first day.

Nvidia is fully committed to developing AI Agent

To help companies develop more complex and intelligent agents, NVIDIA developed a dedicated model, Llama Nemotron, based on Llama-3.1.

Llama Nemotron has three models: Nano, Super and Ultra, each optimized for different deployment needs. The Nano model is suitable for personal computers and edge devices, such as mobile phones, tablets, laptops, etc.;

The Super model provides the best accuracy and throughput on a single GPU; the Ultra model provides maximum agent accuracy on a multi-GPU server.

In addition, NVIDIA enhanced these models in the post-training stage to improve the capabilities of multi-step mathematical calculations, encoding, reasoning, and complex decision-making, with an accuracy improvement of 20% over the basic model and a reasoning speed optimization of 5 times over other leading open source reasoning models . This enables agents developed based on this model to handle complex automated tasks with higher accuracy and enhance decision-making capabilities.

Microsoft is currently integrating the Llama Nemotron model into Azure AI Foundry to provide customers with services such as Azure AI Agent Service for Microsoft 365.

Accenture will provide the Llama Nemotron model on the AI ​​Refinery platform to help customers develop and deploy customized agents. SAP will enhance its Business AI and Joule AI platforms with the Llama Nemotron model to develop customized agents.

Deloitte will incorporate Llama Nemotron into the newly launched Zora AI Intelligent Agent Platform, which supports and simulates agents, giving them deep automation capabilities and industry-specific business knowledge, and enabling them to make decisions and act like humans.

At the same time, NVIDIA has also open-sourced a framework dedicated to developing, optimizing, and integrating agents, AgentIQ , which integrates NVIDIA NeMo Retrieve and supports multimodal information retrieval. It can process various types of data such as text, images, and videos, helping agents combine different types of information to make more comprehensive decisions.

AgentIQ provides a rich set of tools and interfaces that can connect different agents to data sources, support real-time data acquisition and processing, and ensure that agents can reason and make decisions based on the latest information.

AgentIQ also supports building continuous learning capabilities. By combining with NVIDIA NeMo microservices, agents can continuously learn autonomously from feedback generated by humans and AI to optimize their performance and decision-making capabilities.

Open source address: https://github.com/NVIDIA/AgentIQ

Today, NVIDIA and Oracle have also reached in-depth technical cooperation in the field of Agents. The integration of Oracle Cloud Platform and NVIDIA AI Enterprise Software Platform provides more than 160 native AI tools and more than 100 NVIDIA NIM microservices to help enterprises quickly build Agents for complex automation businesses.

In terms of deployment, NVIDIA AI enterprise software not only supports OCI bare metal instances, but can also be deployed in Kubernetes clusters through the OCI Kubernetes engine.

It also provides a code-free deployment solution, allowing customers to quickly run AI workloads without manually configuring infrastructure or selecting a software stack. This provides an efficient solution for NVIDIA GPUs, NIM microservices, and pre-packaged observability tools, helping enterprises reduce Agents from weeks to minutes.

Currently, Pipefy has been using the collaborative cloud platform of Oracle and NVIDIA to perform automatic document preprocessing and image processing, significantly reducing deployment time and saving a lot of operating resources.

Release the most powerful AI factory

In order to accelerate the reasoning efficiency of agents and large models, NVIDIA released the latest version of NVIDIA Blackwell Ultra, an acceleration platform dedicated to AI agents and physical AI.

Blackwell Ultra builds on the groundbreaking Blackwell architecture introduced a year ago and includes the NVIDIA GB300 NVL72 rack-scale solution and the NVIDIA HGX B300 NVL16 system.

The GB300 NVL72 delivers 1.5 times the AI ​​performance of the previous generation GB200 NVL72 and increases Blackwell’s AI factory revenue opportunity by 50 times compared to an AI factory built on NVIDIA Hopper.

The NVIDIA GB300 NVL72 connects 72 Blackwell Ultra GPUs and 36 Arm Neoverse-based NVIDIA Grace™ CPUs in a rack-scale design to act as one giant GPU for scaling while testing.

With the GB300 NVL72, AI models can access the platform’s enhanced computing power, explore different solutions to problems, and break down complex requests into multiple steps, generating higher-quality responses.

Blackwell Ultra is designed for AI agents that use sophisticated reasoning and iterative planning to autonomously solve complex, multi-step problems. Agent systems not only follow instructions, but also reason, plan, and take actions to achieve specific goals.

Physical AI helps companies generate synthetic, realistic videos in real time and train robots and self-driving cars at scale.

Blackwell Ultra systems can be seamlessly integrated into NVIDIA Spectrum-X™ Ethernet and NVIDIA Quantum-X800 InfiniBand platforms, and each GPU in the system can get 800 Gb/s of data throughput through the NVIDIA ConnectX®-8 SuperNIC. This provides industry-leading remote direct memory access capabilities, enabling AI factories and cloud data centers to process AI inference models without bottlenecks.

In addition, the Blackwell Ultra system is equipped with NVIDIA BlueField®-3 DPUs, which support multi-tenant networking, GPU computing elasticity, accelerated data access and real-time network security threat detection.

Nvidia's partners will launch products based on Blackwell Ultra starting in the second half of 2025. Companies such as Cisco, Dell Technologies, HPE, Lenovo and Supermicro will offer a range of servers based on Blackwell Ultra products, as will companies such as Aivres, ASRock Rack, Asus, Eviden, Foxconn, Gigabyte, Inventec, Quanta Cloud Technology, Wistron and VIA.

Cloud service providers Amazon Web Services, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure, as well as GPU cloud providers CoreWeave, Crusoe, Lambda, Nebius, Nscale, Yotta and YTL will be the first to offer Blackwell Ultra.

"AI has made great strides, and agents need higher computing performance," said Jen-Hsun Huang, founder and CEO of NVIDIA . "So, we designed Blackwell Ultra, a versatile platform that can easily and efficiently perform pre-training, post-training, and inference."

The source of this article is NVIDIA's official website. If there is any infringement, please contact us to delete it.

END