Nemotron 3 Super

Nemotron 3 Super is a 120B hybrid Mamba-Transformer mixture-of-experts model, with 12B active parameters per forward pass. It is designed to coordinate multi-agent systems: routing work, planning, and bringing results together. It uses four experts per pass and has a 262,144-token context window with tool calling. For routine, high-volume tasks, NVIDIA recommends pairing it with a smaller model such as Nemotron 3 Nano. The weights, training data, and recipes are open source.