NVIDIA's Mixture-of-Experts model with 120B total parameters and 12B active, optimized for efficient inference with strong reasoning capabilities.