Nemotron 3 Ultra

Designed for Agents: More than just a chat model, it is purpose-built for complex agentic workflows that require multi-step reasoning, tool calling, planning, and self-correction.

Powerful MoE Architecture: With 550B total parameters, only 55B are activated per inference (Mixture of Experts), delivering high performance while maintaining efficiency.

Ultra-Long Context Window: Supports up to 1M tokens, enabling one-pass processing of lengthy reports, large codebases, and other extensive documents.

Superior Inference Efficiency: Through innovations like NVFP4 precision and Multi-Token Prediction (MTP), inference throughput is up to 5× higher than comparable models, reducing the cost of complex agent tasks by up to 30%.

Broad Ecosystem Compatibility: Already integrated with leading agent development frameworks (e.g., OpenHands, LangChain Deep Agents), making it easy for developers to quickly integrate and deploy.

1. AI Agent Developers
Teams building complex agentic applications for coding, research, and enterprise automation, requiring multi-step reasoning and tool-calling capabilities.

2. Enterprise IT and Industry Users
Applicable to software, cybersecurity, legal, R&D, and more, handling long-running tasks such as large codebases, vulnerability fixes, and massive text analysis.

3. Organizations Prioritizing Data Sovereignty and Cost
Open-source with on-premise deployment options, meeting privacy requirements for government and finance. Inference speed increases by up to 5× while costs drop by 30%, balancing performance and efficiency.