News Summary:<\/p>\n
\n\tThe Nemotron 3 family of open models \u2014 in\u00a0Nano, Super and Ultra sizes \u2014 introduces the most efficient family of open models with leading accuracy for building\u00a0agentic\u00a0AI applications.
\n\tNemotron 3 Nano delivers 4x higher throughput than Nemotron 2 Nano and delivers the most tokens per second for multi-agent systems at scale through a breakthrough hybrid mixture-of-experts architecture.
\n\tNemotron\u00a0achieves superior accuracy from advanced reinforcement learning techniques with concurrent multi-environment post-training at scale.
\n\tNVIDIA is the first to release a collection of state-of-the-art open models, training datasets and reinforcement learning environments and libraries for building highly accurate, efficient, specialized AI agents.
<\/p>\n
\u00a0<\/p>\n
NVIDIA today announced the NVIDIA Nemotron\u2122 3 family of open models, data and libraries designed to power transparent, efficient and specialized agentic AI development across industries.<\/p>\n
The Nemotron 3 models \u2014 with Nano, Super and Ultra sizes \u2014 introduce a breakthrough hybrid latent mixture-of-experts (MoE)<\/a> architecture that helps developers build and deploy reliable multi-agent systems at scale.<\/p>\n As organizations shift from single-model chatbots to collaborative multi-agent AI systems, developers face mounting challenges, including communication overhead, context drift and high inference costs. In addition, developers require transparency to trust the models that will automate their complex workflows. Nemotron 3 directly addresses these challenges, delivering the performance and openness customers need to build specialized, agentic AI.<\/p>\n \u201cOpen innovation is the foundation of AI progress,\u201d said Jensen Huang, founder and CEO of NVIDIA. \u201cWith Nemotron, we\u2019re transforming advanced AI into an open platform that gives developers the transparency and efficiency they need to build agentic systems at scale.\u201d<\/p>\n NVIDIA Nemotron supports NVIDIA\u2019s broader sovereign AI efforts, with organizations from Europe<\/a> to South Korea<\/a> adopting open, transparent and efficient models that allow them to build AI systems aligned to their own data, regulations and values.<\/p>\n Early adopters, including Accenture, Cadence, CrowdStrike, Cursor, Deloitte, EY, Oracle Cloud Infrastructure, Palantir, Perplexity, ServiceNow, Siemens, Synopsys and Zoom, are integrating models from the Nemotron family to power AI workflows across manufacturing, cybersecurity, software development, media, communications and other industries.<\/p>\n \u201cNVIDIA and ServiceNow have been shaping the future of AI for years, and the best is yet to come,\u201d Bill McDermott, chairman and CEO of ServiceNow. \u201cToday, we\u2019re taking a major step forward in empowering leaders across all industries to fast-track their agentic AI strategy. ServiceNow\u2019s intelligent workflow automation combined with NVIDIA Nemotron 3 will continue to define the standard with unmatched efficiency, speed and accuracy.\u201d<\/p>\n As multi-agent AI systems expand, developers are increasingly relying on proprietary models for state-of-the-art reasoning while using more efficient and customizable open models to drive down costs. Routing tasks between frontier-level models and Nemotron in a single workflow gives agents the most intelligence while optimizing tokenomics<\/a>.<\/p>\n “Perplexity is built on the idea that human curiosity will be amplified by accurate AI built into exceptional tools, like AI assistants,” said Aravind Srinivas, CEO of Perplexity. \u201cWith 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 \u2014 ensuring our AI assistants operate with exceptional speed, efficiency and scale.\u201d<\/p>\n The open Nemotron 3 models enable startups to build and iterate faster on AI agents and accelerate innovation from prototype to enterprise deployment. General Catalyst and Mayfield\u2019s portfolio companies are exploring Nemotron 3 to build AI teammates that support human-AI collaboration.<\/p>\n \u201cNVIDIA\u2019s open model stack and the NVIDIA Inception program give early-stage companies the models, tools and a cost-effective infrastructure to experiment, differentiate and scale fast,\u201d said Navin Chaddha, managing partner at Mayfield. \u201cNemotron 3 gives founders a running start on building agentic AI applications and AI teammates, and helps them tap into NVIDIA\u2019s massive installed base.\u201d<\/p>\n Nemotron 3 Reinvents Multi-Agent AI With Efficiency and Accuracy
Available today, Nemotron 3 Nano is the 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 unique hybrid MoE architecture to deliver gains in efficiency and scalability.<\/p>\n This design achieves up to 4x 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 remembers more, making it more accurate and better capable of connecting information over long, multistep tasks.<\/p>\n Artificial Analysis, an independent organization that benchmarks AI, ranked the model as the most open and efficient among models of the same size, with leading accuracy.<\/p>\n Nemotron 3 Super excels at applications that require many collaborating agents to achieve complex tasks with low latency. Nemotron 3 Ultra serves as an advanced reasoning engine for AI workflows that demand deep research and strategic planning.<\/p>\n Nemotron 3 Super and Ultra use NVIDIA\u2019s ultraefficient 4-bit NVFP4 training format on the NVIDIA Blackwell architecture, significantly cutting memory requirements and speeding up training. This efficiency allows larger models to be trained on existing infrastructure without compromising accuracy relative to higher-precision formats.<\/p>\n With the Nemotron 3 family of models, developers can choose the open model that is right-sized for their specific workloads, scaling from dozens to hundreds of agents while benefiting from faster, more accurate long-horizon reasoning for complex workflows.<\/p>\n New Open Tools and Data for AI Agent Customization Three trillion tokens of new Nemotron pretraining<\/a>, post-training<\/a> and reinforcement learning<\/a> datasets supply the rich reasoning, coding and multistep workflow examples needed to create highly capable, domain-specialized agents. The Nemotron Agentic Safety Dataset<\/a> provides real-world telemetry to help teams evaluate and strengthen the safety of complex agent systems.<\/p>\n To accelerate development, NVIDIA released the NeMo Gym<\/a> and NeMo RL<\/a> open-source libraries, which provide the training environments and post-training foundation for Nemotron models, along with NeMo Evaluator to validate model safety and performance. All tools and datasets are now available on GitHub and Hugging Face.<\/p>\n Nemotron 3 is supported by LM Studio<\/a>, llama.cpp, SGLang<\/a> and vLLM<\/a>. In addition, Prime Intellect and Unsloth<\/a> are integrating NeMo Gym\u2019s ready-to-use training environments directly into their workflows, giving teams faster, easier access to powerful reinforcement learning training.<\/p>\n Get Started With NVIDIA Open Models Nemotron is offered on enterprise AI and data infrastructure platforms, including Couchbase, DataRobot, H2O.ai, JFrog, Lambda and UiPath. For customers on public clouds, Nemotron 3 Nano will be available on AWS via Amazon Bedrock (serverless) as well as supported on Google Cloud, CoreWeave, Crusoe, Microsoft Foundry, Nebius<\/a>, Nscale and Yotta soon.<\/p>\n
\nThe Nemotron 3 family of MoE models<\/a> includes three sizes:<\/p>\n
\n\tNemotron 3 Nano, a small, 30-billion-parameter model that activates up to 3 billion parameters at a time for targeted, highly efficient tasks.
\n\tNemotron 3 Super, a high-accuracy reasoning model with approximately 100 billion parameters and up to 10 billion active per token, for multi-agent applications.
\n\tNemotron 3 Ultra, a large reasoning engine with about 500 billion parameters and up to 50 billion active per token, for complex AI applications.
<\/p>\n
\nNVIDIA also released a collection of training datasets and state-of-the-art reinforcement learning libraries available to anyone building specialized AI agents.<\/p>\n
\nNemotron 3 Nano is available today on Hugging Face<\/a> and through inference service providers including Baseten<\/a>, DeepInfra<\/a>, Fireworks<\/a>, FriendliAI<\/a>, OpenRouter<\/a> and Together AI<\/a>.<\/p>\n