Nemotron is a family of open-weight models designed specifically for agentic AI tasks, integrated deeply throughout the NVIDIA Agent Toolkit. Nemotron models deliver the reasoning capability, efficiency, and specialization required for production agents without proprietary model dependencies. The Nemotron 3 family (Nano, Super, Ultra sizes) uses a hybrid Mamba-Transformer mixture-of-experts (MoE) architecture with 1M-token context window, optimized for complex, high-throughput agent reasoning.
Model capabilities span agentic task domains: Nemotron for general reasoning, Nemotron Vision for visual understanding, Nemotron RAG for retrieval-augmented generation, Nemotron Guardrail for safety and compliance, and Nemotron Speech for voice interaction. All are open-weight with open training data and recipes, enabling organizations to fine-tune models on proprietary domain data and maintain full control over agent reasoning.
Within the toolkit, Nemotron models power the AI-Q Blueprint's research agents, providing cost-effective reasoning while frontier models handle orchestration. Developers can run Nemotron on-premises using vLLM, SGLang, Ollama, or llama.cpp, or access them as NVIDIA NIM microservices for managed inference. For enterprise deployments with Zilliz Cloud, Nemotron models deployed alongside managed vector databases enable fully-integrated agentic RAG infrastructure—agents retrieve context from Zilliz, reason using Nemotron, and produce assured answers grounded in enterprise knowledge.