NVIDIA ships Nemotron 3 Ultra — fully open 550B reasoning model built for long-running agents

NVIDIA released Nemotron 3 Ultra (4 Jun), its largest open-weight reasoning model, explicitly tuned for multi-step, long-horizon agentic work rather than chat. A real build-vs-buy lever: a frontier-class, self-hostable reasoning brain for governed on-prem pipelines without per-token API exposure.