NVIDIA has launched the Nemotron 3 family, featuring three model sizes: Nano, Super, and Ultra. These models utilize a hybrid mixture-of-experts architecture aimed at improving efficiency and accuracy for multi-agent AI systems. The Nemotron 3 Nano offers a significant performance boost, with a reported 4x increase in throughput compared to its predecessor, Nemotron 2 Nano. With 30 billion parameters, it activates up to 3 billion at once, making it well-suited for tasks like software debugging and content summarization. Its design allows for a 1-million-token context window, enhancing its ability to handle complex, multi-step tasks. The Super model, with around 100 billion parameters, and the Ultra model, boasting about 500 billion parameters, cater to more demanding applications requiring collaboration among multiple agents. The Super model excels in low-latency scenarios, while the Ultra serves as a powerful reasoning engine for deep research and strategic tasks. Both models leverage NVIDIA’s NVFP4 training format, which cuts memory requirements and speeds up training, allowing for larger models without sacrificing accuracy. NVIDIA is also providing extensive resources for developers, including three trillion tokens of training data and open-source libraries like NeMo Gym for reinforcement learning. This collection aims to help teams create specialized AI agents with a focus on safety and performance. Early adopters such as Accenture and Oracle are already integrating these models into their workflows across various sectors, from manufacturing to cybersecurity.