Nvidia released Nemotron 3.5 Lightning yesterday, a 30B open mixture-of-experts model with 3B active parameters. It also released NeMo Switchyard, an open-source router that decides which model should handle each step of an agent workflow.
The model is useful. The router is the more interesting part.
Most agent stacks eventually hit the same ugly fork. You can send everything to the best model and pay for it, or you can write a little routing layer that sends easy steps to cheaper models. That little routing layer then starts growing teeth. It needs policies. It needs fallback logic. It needs evals. It needs to know when a cheap model is being confidently wrong. It needs enough logging that you can explain why a run cost $2.40 instead of $0.18.
At some point the agent is no longer one model. It is a small dispatch system with a model attached.
That is what Nvidia is trying to productize with Switchyard.
