NVIDIA has announced the Nemotron™ 3 family of open models, along with new datasets and open-source tools designed to support transparent, efficient, and specialized agentic AI development across industries. The release reflects a broader shift from single-model chatbots toward multi-agent AI systems that collaborate to perform complex tasks at scale.

The Nemotron 3 family introduces a new hybrid latent mixture-of-experts (MoE) architecture, aimed at reducing inference costs while improving reliability and transparency—two challenges developers increasingly face as AI agents take on more autonomous roles in enterprise workflows.

As organizations move beyond single AI assistants toward collaborative agent systems, developers face growing challenges such as communication overhead, context drift, and rising compute costs. At the same time, trust and transparency are becoming critical as agents automate more complex workflows.

NVIDIA positioned Nemotron 3 as a response to these pressures, combining open access with architectural efficiency to support agentic systems that can scale without relying entirely on expensive proprietary models.

“Open innovation is the foundation of AI progress,” said Jensen Huang, founder and CEO of NVIDIA. “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.”

The Nemotron 3 lineup includes three model sizes, each designed for different agentic workloads:

While the Nemotron 3 models vary in size and capability, much of their efficiency gains stem from a shared hybrid mixture-of-experts (MoE) architecture.

Nemotron 3 Nano, available today, is NVIDIA’s most compute-cost-efficient model, optimized for tasks such as software debugging, content summarization, AI assistant workflows, and information retrieval at low inference costs. The model uses a hybrid MoE architecture designed to improve both efficiency and scalability for agentic AI workloads.

According to NVIDIA, this design delivers up to 4× higher token throughput compared with Nemotron 2 Nano and reduces reasoning-token generation by up to 60%, significantly lowering inference costs. With a 1-million-token context window, Nemotron 3 Nano can maintain longer context across multistep workflows, improving accuracy and its ability to connect information over extended tasks.

Artificial Analysis, an independent AI benchmarking organization, ranked Nemotron 3 Nano among the most open and efficient models in its size class, with strong accuracy benchmarks.

For Nemotron 3 Super and Ultra, NVIDIA uses its ultraefficient 4-bit NVFP4 training format on the Blackwell architecture, reducing memory requirements while maintaining accuracy. This enables larger models to be trained and deployed on existing infrastructure without the overhead typically associated with higher-precision formats.

Nemotron 3 Super is designed for applications that require many collaborating agents to complete complex tasks with low latency, while Nemotron 3 Ultra serves as a more advanced reasoning engine for AI workflows that demand deeper research and strategic planning. Together, the Nemotron 3 family allows developers to select open models that are right-sized for their workloads—scaling from dozens to hundreds of agents while supporting faster, more accurate long-horizon reasoning across complex workflows.

As multi-agent AI systems become more complex, developers are increasingly combining proprietary frontier models with more efficient and customizable open models to balance performance and cost. NVIDIA said routing tasks between frontier-level models and Nemotron within a single workflow allows agents to use advanced reasoning where it matters most, while relying on lower-cost open models for coordination and execution—helping teams optimize both intelligence and token usage.

Early adopters of Nemotron 3 span industries including manufacturing, cybersecurity, software development, media, and communications. Companies integrating Nemotron models include Accenture, Cadence, CrowdStrike, Cursor, Deloitte, EY, Oracle Cloud Infrastructure, Palantir, Perplexity, ServiceNow, Siemens, Synopsys, and Zoom.

“ServiceNow’s intelligent workflow automation combined with NVIDIA Nemotron 3 will continue to define the standard with unmatched efficiency, speed, and accuracy,” said Bill McDermott, chairman and CEO of ServiceNow.

Perplexity also highlighted how Nemotron 3 fits into hybrid AI strategies that route tasks between open and proprietary models.

“With our agent router, we can direct workloads to the best fine-tuned open models, like Nemotron 3 Ultra, or leverage leading proprietary models when tasks benefit from their unique capabilities—ensuring our AI assistants operate with exceptional speed, efficiency, and scale,” said Aravind Srinivas, CEO of Perplexity.

The open Nemotron 3 models are also aimed at accelerating innovation for startups and early-stage teams building AI agents. NVIDIA said the models allow developers to move faster from prototype to enterprise deployment, supporting the creation of AI teammates designed for human–AI collaboration.

Venture-backed companies across the portfolios of General Catalyst, Mayfield, and Sierra Ventures are already exploring Nemotron 3 as part of their agentic AI development efforts.

“Nemotron 3 gives founders a running start on building agentic AI applications and AI teammates, and helps them tap into NVIDIA’s massive installed base,” said Navin Chaddha, managing partner at Mayfield.

Alongside the Nemotron 3 models, NVIDIA released a broad collection of training datasets and reinforcement learning tools designed to support teams building specialized agentic AI systems.

The release includes three trillion tokens of new Nemotron datasets spanning pretraining, post-training, and reinforcement learning, providing reasoning, coding, and multistep workflow examples needed to create highly capable, domain-specific agents. NVIDIA also introduced the Nemotron Agentic Safety Dataset, which uses real-world telemetry to help teams evaluate and strengthen the safety of complex multi-agent systems.

To accelerate development and customization, NVIDIA released the open-source NeMo Gym and NeMo RL libraries, which provide training environments and post-training foundations for Nemotron models, along with NeMo Evaluator to validate model performance and safety. All datasets and tools are available on GitHub and Hugging Face.

Nemotron 3 is supported by widely used inference and development frameworks, including LM Studio, llama.cpp, SGLang, and vLLM. In addition, Prime Intellect and Unsloth are integrating NeMo Gym’s ready-to-use training environments directly into their workflows, giving teams faster and easier access to reinforcement learning training for agentic AI.

Nemotron 3 Nano is available today on Hugging Face and through inference providers such as Baseten, DeepInfra, Fireworks, FriendliAI, OpenRouter, and Together AI. It is also offered as an NVIDIA NIM™ microservice for secure, scalable deployment.

For customers on public clouds, Nemotron 3 Nano is available on AWS via Amazon Bedrock (serverless) and is also supported on Google Cloud, Microsoft Foundry, CoreWeave, Crusoe, Nebius, Nscale, and Yotta, with broader availability expanding over time.

Beyond cloud deployment, Nemotron is also being integrated across a range of enterprise AI and data infrastructure platforms, including Couchbase, DataRobot, H2O.ai, JFrog, Lambda, and UiPath, allowing organizations to deploy the models within existing enterprise workflows.

Nemotron 3 Super and Ultra are expected to become available in the first half of 2026.

Q: What is Nemotron 3 designed for?
A: Nemotron 3 is designed to power multi-agent AI systems, enabling multiple AI agents to collaborate efficiently across complex workflows.

Q: Why does NVIDIA emphasize open models?
A: Open models allow organizations to customize AI systems, reduce costs, and align deployments with their own data, regulations, and values.

Q: How does Nemotron 3 reduce inference costs?
A: Its hybrid MoE architecture activates only a subset of parameters per task, improving efficiency and lowering compute requirements.

Q: Who is adopting Nemotron 3 today?
A: Early adopters include enterprises, cloud providers, and AI platforms such as Accenture, ServiceNow, Perplexity, and Siemens.

As AI systems evolve from single assistants into networks of collaborating agents, the cost and complexity of running those systems is becoming a real constraint. Without more efficient and transparent models, agentic AI risks becoming accessible only to the largest companies with the deepest compute budgets.

NVIDIA’s Nemotron 3 highlights a different path: one where open, right-sized models handle much of the reasoning and coordination work, while frontier models are used selectively. This approach lowers costs, reduces infrastructure strain, and gives developers more control over how AI agents behave within complex workflows.

For enterprises, this matters because agentic AI is moving closer to automating real operational decisions—not just generating responses. Open models make it easier to inspect, customize, and govern those systems, especially in regulated industries or sovereign AI initiatives where alignment with local data and values is critical.

For startups and developers, Nemotron 3 represents a shift in who gets to build agentic AI at scale. By lowering inference costs and offering open tools for customization and safety evaluation, NVIDIA is helping broaden access to multi-agent systems that were previously out of reach.

As AI agents take on more responsibility across software, cybersecurity, manufacturing, and knowledge work, the future of agentic AI will depend less on raw model size—and more on efficiency, openness, and the ability to keep humans meaningfully in control.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.