The Nemotron 3 lineup – comprising Nano, Tremendous, and Extremely – delivers main efficiency for multi-agent AI techniques, combining superior reasoning, conversational, and collaborative capabilities. The fashions leverage a hybrid Mamba-Transformer mixture-of-experts (MoE) structure, offering best-in-class inference throughput whereas supporting context lengths of as much as 1 million tokens.
Nemotron 3 Nano, the smallest mannequin, is optimized for cost-efficient inference and duties equivalent to software program debugging, content material summarization, AI assistant workflows, and knowledge retrieval. Regardless of possessing 30 billion whole parameters, it intelligently prompts solely about 3 billion per token. With a novel hybrid MoE design, Nano achieves as much as 4× larger token throughput than its predecessor and reduces reasoning-token era by 60%, all whereas sustaining superior accuracy. Early benchmarks present Nano outperforming comparable open fashions like GPT-OSS-20B and Qwen3-30B on reasoning and long-context duties.
Nemotron 3 Tremendous and Extremely lengthen these capabilities for high-volume collaborative brokers and sophisticated AI functions, incorporating improvements equivalent to latent MoE, a hardware-aware professional design that will increase mannequin high quality with out sacrificing effectivity, and multi-token prediction (MTP), which boosts long-form textual content era and multi-step reasoning. Each bigger fashions are skilled utilizing NVIDIA’s NVFP4 format, enabling sooner coaching and decreased reminiscence necessities.
All Nemotron 3 fashions are post-trained utilizing multi-environment reinforcement studying (RL), enabling them to deal with duties spanning mathematical and scientific reasoning, aggressive coding, instruction following, software program engineering, chat, and multi-agent device use. The fashions additionally assist granular reasoning funds management at inference time, permitting builders to fine-tune computational sources whereas sustaining accuracy.
NVIDIA has additionally launched a complete suite of datasets, coaching libraries, and analysis instruments, together with over three trillion tokens of pretraining and reinforcement studying knowledge, the NeMo Health club and NeMo RL open-source libraries, and the Nemotron Agentic Security Dataset for real-world security analysis.
The Nemotron 3 household is designed to empower builders, startups, and enterprises to construct specialised AI brokers transparently and effectively. Nano is obtainable right this moment by Hugging Face, NVIDIA NIM microservices, and main cloud and AI platforms together with AWS, Google Cloud, and Microsoft Foundry. Tremendous and Extremely are anticipated to launch within the first half of 2026.
Early adopters equivalent to Accenture, ServiceNow, Perplexity, and Palantir are already integrating Nemotron 3 fashions into AI workflows for manufacturing, cybersecurity, software program growth, media, and enterprise operations.
With Nemotron 3, NVIDIA is engaged on a brand new normal for environment friendly, correct, and open AI fashions. This can enable builders to scale agentic AI functions from prototype to enterprise deployment whereas sustaining transparency, cost-efficiency, and state-of-the-art efficiency.
