Models
NVIDIA pairs a 3B-active MoE model with agent workload routing
NVIDIA released Nemotron 3.5 Lightning, a 30B mixture-of-experts model that uses 3B parameters at a time. NeMo Switchyard can send each agent task to a model based on quality and cost instead of using one large model for everything.
Why it matters
An agent does not need the same model for every step. A strong model can plan while smaller models handle repeated tool calls and checks. The router now becomes an important part of the infrastructure because it affects quality, speed, cost, privacy, and failures.
Tejo's take
Test the model and router together. Measure wrong routing decisions and cost. Define what happens when a model fails, and keep credentials and logging outside the router. NVIDIA reported the performance numbers, so wait for independent tests too.