NVIDIA's Nemotron 3 family introduces a new approach to agentic AI systems, emphasizing collections of cooperating agents that work together over extended contexts. The key innovations include a hybrid architecture combining Mamba layers, Transformer layers, and a mixture-of-experts (MoE) system. This design allows for efficient processing of long sequences while maintaining high accuracy in reasoning. The models support a 1M-token context length, which is significant for tasks requiring deep document reasoning and sustained interactions, enabling agents to manage extensive data sets without fragmenting information. Nemotron 3 is built for real-world applications, using multi-environment reinforcement learning via NeMo Gym. This training method ensures the models can handle complex workflows, not just single-turn tasks. The open training pipeline makes it easy for developers to customize and adapt the models for specific needs. The initial model, Nemotron 3 Nano, is ready for deployment now, optimized for various NVIDIA GPUs, while the more advanced Super and Ultra models will be released in the first half of 2026. Upcoming features in Super and Ultra include latent MoE for better expert utilization and multi-token prediction (MTP) for faster output generation. These enhancements aim to improve the efficiency and depth of reasoning capabilities, making the models more effective for planning and code generation tasks. NVIDIA is also committed to open models, providing access to model weights, training datasets, and detailed recipes for reproducibility and customization.