Nemotron 3.5 Super VL uses a Mixture-of-Experts architecture where only 12 billion of its 120 billion total parameters are active per token