NVIDIA announced the Nemotron™ 3 family of open models, designed to accelerate the development of specialized, agentic AI across industries. Available in Nano, Super, and Ultra sizes, Nemotron 3 offers high efficiency and accuracy. Its hybrid latent mixture-of-experts (MoE) architecture supports scalable multi-agent AI systems while ensuring transparency and control.
As organizations shift from single-model chatbots to collaborative AI workflows, developers encounter challenges. These include context drift, communication overhead, and high inference costs. Nemotron 3 addresses these issues by combining advanced reinforcement learning with multi-environment post-training, providing reliable reasoning for complex, multi-step tasks.
“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.”
Early adopters including Accenture, ServiceNow, CrowdStrike, Deloitte, and Perplexity are integrating Nemotron 3 into workflows spanning manufacturing, cybersecurity, media, and software development. Bill McDermott, CEO of ServiceNow, noted, “ServiceNow’s intelligent workflow automation combined with NVIDIA Nemotron 3 will continue to define the standard with unmatched efficiency, speed and accuracy.”
Also Read: Cohere Unveils Rerank4, a More Powerful Search and Retrieval Model for Enterprise AI
The Nemotron 3 family includes:
Nano: 30B parameters, up to 3B active per token, optimized for high-throughput tasks such as debugging, summarization, and AI assistants, achieving 4× higher token throughput than Nemotron 2 Nano.
Super: 100B parameters, up to 10B active per token, designed for low-latency multi-agent collaboration.
Ultra: 500B parameters, up to 50B active per token, suited for deep reasoning and complex AI workflows.
All models leverage NVIDIA’s ultra-efficient 4-bit NVFP4 training format on the Blackwell architecture, reducing memory requirements while maintaining high accuracy.
In addition, NVIDIA released NeMo Gym and NeMo RL libraries, 3 trillion tokens of pretraining datasets, and agentic safety datasets to accelerate AI agent development. Nemotron 3 Nano is available today via Hugging Face, enterprise AI platforms, and cloud providers including AWS, with Super and Ultra models expected in H1 2026.
Nemotron 3 empowers developers—from startups to enterprises—to scale efficient, accurate, and transparent multi-agent AI, driving innovation from prototype to production.