Nvidia Switchyard Cuts AI Task Costs by Two-Thirds

Nvidia shipped two products on Tuesday aimed at a problem every company running AI agents eventually runs into: send every task to a frontier model and costs balloon, but build your own logic to route cheap tasks elsewhere and you've created a second engineering job. The first product is Nemotron 3.5 Lightning, a 30-billion-parameter open mixture-of-experts model built for high-volume agent work. The second is NeMo Switchyard, an open-source library that decides, step by step, which model should handle which part of a workflow.

Nvidia says Lightning runs up to four times faster than comparable models and finishes agentic tasks about 30% quicker than Qwen3.6-35B at the same accuracy. Paired with Switchyard, the company claims the setup matches frontier-level task completion while dropping benchmark costs to roughly a third of running Opus 4.8 by itself.

The release lands in the middle of the busiest stretch of open-weight competition the industry has had in months, with Chinese labs and Meta all pushing out models that blur the line between open and frontier. Nvidia's numbers, drawn from its own testing, are what the company is putting forward as evidence.