AI MODELS

5 min read

NVIDIA's Nemotron 3 Super combines latent mixture of experts with hybrid Mamba architecture - 120B total parameters, 12B active per token, 1M context, and up to 4x more experts at the same cost.