Nemotron 3 Super 120B A12B is a sparse mixture-of-experts model that punches above its weight by activating only 12 billion parameters at a time despite having 120 billion total. This makes it notably efficient at inference — you get large-model reasoning without paying full large-model compute costs. It handles general language tasks competently, though its MoE architecture means behavior can occasionally feel less consistent than dense models of comparable active parameter counts.