NVIDIA opens a 30B model router stack

NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model, alongside the open-source NeMo Switchyard routing library. NVIDIA’s launch note reports up to 4× faster output and says Switchyard’s internal tests cut task-completion cost to nearly one-third of Opus 4.8; the public routing code lets teams route requests across open, proprietary, and NVIDIA models. The stack is a routing signal, not a universal benchmark win: teams evaluating agent harness cost and reliability should reproduce the quality-cost trade-off on their own traces before replacing a single frontier default.