NVIDIA has taken a major step toward the next evolution of artificial intelligence with the announcement of NVIDIA Nemotron 3, a new family of open AI models, datasets, and libraries designed to power transparent, cost-efficient, and scalable multi-agent AI systems. The launch signals NVIDIA’s growing focus on agentic AI—systems where multiple AI models collaborate to solve complex, real-world tasks across industries.
A New Open Platform for Agentic AI
Unveiled on Monday, the Nemotron 3 family reflects NVIDIA’s commitment to open innovation and developer flexibility. According to the company, these models are designed to help organisations transition from single AI chatbots to coordinated, multi-agent workflows that can reason, plan, and execute tasks more efficiently.
“Open innovation is the foundation of AI progress,” said NVIDIA founder and CEO Jensen Huang. “With Nemotron, we’re transforming advanced AI into an open platform that gives developers the transparency and efficiency they need to build agentic systems at scale.”
At the core of Nemotron 3 is a hybrid latent mixture-of-experts (MoE) architecture, engineered to reduce inference costs, minimise context drift, and improve coordination among multiple AI agents—three challenges that have historically limited large-scale agentic AI deployments.
Three Models, Built for Different AI Needs
The Nemotron 3 lineup includes Nano, Super, and Ultra, each tailored to different levels of performance, cost, and reasoning depth.
Nemotron 3 Nano: Fast, Efficient, and Available Now
The first model to launch, Nemotron 3 Nano, is available immediately. It is a 30-billion-parameter model that dynamically activates up to 3 billion parameters per task, making it highly efficient for everyday AI workloads.
Optimised for low-cost inference, Nano is ideal for applications such as:
- Software debugging
- Summarisation
- AI assistants
- General reasoning tasks
NVIDIA reports that Nemotron 3 Nano delivers up to four times higher token throughput than its predecessor, Nemotron 2 Nano, while cutting reasoning token generation by as much as 60%—a significant gain for enterprises focused on performance and cost control.
The model is available on Hugging Face and through leading inference providers, including Baseten, DeepInfra, Fireworks, FriendliAI, OpenRouter, and Together AI. It is also offered as an NVIDIA NIM microservice, enabling seamless deployment on NVIDIA-accelerated infrastructure. Additionally, Nemotron 3 Nano will be accessible via Amazon Bedrock on AWS, with support for multiple cloud platforms scheduled to roll out in the coming months.
Nemotron 3 Super and Ultra: Designed for Advanced Agentic Systems
For more demanding workloads, NVIDIA introduced Nemotron 3 Super and Nemotron 3 Ultra, both expected to become available in the first half of 2026.
- Nemotron 3 Super (~100 billion parameters) is built for low-latency multi-agent applications, where rapid coordination between agents is critical.
- Nemotron 3 Ultra (~500 billion parameters) targets deep reasoning, complex decision-making, and long-horizon planning, making it suitable for advanced enterprise and government use cases.
Both models leverage NVIDIA’s 4-bit NVFP4 training format on Blackwell GPUs, dramatically reducing memory requirements while maintaining high performance—an important factor for scaling large AI systems efficiently.
Powering the Shift to Collaborative AI Agents
The launch of Nemotron 3 comes as enterprises increasingly adopt agent-based AI architectures, where multiple models work together across complex workflows instead of relying on a single monolithic system.
NVIDIA says Nemotron 3 enables developers to intelligently route tasks between proprietary frontier models and open Nemotron models within the same workflow. This approach allows organisations to balance reasoning depth, performance, and cost efficiency—using premium models only when necessary and open models where they perform best.
Supporting Sovereign AI and Global Adoption
Nemotron 3 also aligns closely with NVIDIA’s sovereign AI strategy, empowering governments and enterprises to deploy AI systems tailored to local data, regulatory requirements, and policy frameworks. According to NVIDIA, organisations across Europe and South Korea are already adopting the open models to support region-specific AI initiatives.
Strong Enterprise and Partner Ecosystem
A wide range of enterprise leaders and technology partners are integrating Nemotron models into their AI workflows. These include Accenture, Deloitte, EY, Oracle Cloud Infrastructure, Palantir, Perplexity, ServiceNow, Siemens, Synopsys, and Zoom, with applications spanning manufacturing, cybersecurity, software development, and enterprise communications.
Aravind Srinivas, CEO of Perplexity, highlighted the flexibility Nemotron brings to agent routing systems: “We can direct workloads to fine-tuned open models like Nemotron 3 Ultra or use proprietary models when tasks require it.”
Open Datasets and Tools for Agentic AI Development
Alongside the models, NVIDIA released three trillion tokens of pretraining, post-training, and reinforcement learning datasets. This includes a new Agentic Safety Dataset designed to evaluate and improve the behaviour of multi-agent systems.
To further support developers, NVIDIA has also open-sourced key tools, including:
- NeMo Gym
- NeMo RL
- NeMo Evaluator
These tools enable end-to-end training, customisation, and evaluation of agentic AI systems, accelerating innovation while maintaining transparency and safety.
A Defining Moment for Open, Agentic AI
With Nemotron 3, NVIDIA is positioning itself at the centre of the shift toward open, collaborative, and sovereign AI systems. By combining powerful open models, enterprise-grade tooling, and a strong ecosystem of partners, the company is laying the groundwork for the next generation of multi-agent AI—one that is more efficient, adaptable, and globally deployable than ever before.
Read more for more latest articles post at Edelman Appoints Anna Vogt as Global Chief Strategy Officer