General purpose reasoning and chat model trained from scratch by NVIDIA. Contains 30B total parameters with only 3.5B active at a time for low-latency MoE inference