Nvidia bets on cheap specialist models plus a smart router, not bigger single models, for agentic AI

Agentic workloads run continuously and hit models thousands of times per session — for tool calls, alert triage, billing lookups. Routing every one of those through a single frontier reasoning model is both slow and expensive. Nvidia's answer is a two-part release: Nemotron 3.5 Lightning, a 30B mixture-of-experts model tuned for high-volume specialized tasks, and NeMo Switchyard, an open-source router that sends each step of a workflow to whichever model — Nvidia's, open, or proprietary — fits the job's quality/latency/cost profile.

Lightning runs from RTX PCs, DGX Spark/Station and Jetson at the edge up to RTX PRO workstations, data centers and cloud — the same model scales down for local/private use and up for enterprise volume. As with prior Nemotron releases, Nvidia publishes much of the training data and technique (licensing permitting), plus a new dataset, Nemotron-RL-Agentic-Terminal-Pivot, used to post-train the model's coding-agent behavior.