NVIDIA has released Nemotron 3.5 Lightning, a 30B open mixture-of-experts (MoE) model with only 3B active parameters, alongside the NeMo Switchyard model router. The pairing targets the agent execution layer, where Switchyard routes each step to the cheapest capable model — a move that could directly cut inference costs for organizations running AI agents at scale.
The model's efficiency is notable: with 30B total parameters but only 3B active per forward pass, it promises lower latency and cost than dense models of similar capability. For decision-makers, the real value lies in the router: instead of paying premium rates for every agent step, Switchyard dynamically selects the most economical model that can handle the task. This is a practical answer to the cost explosion many enterprises face when deploying multi-step agents.
NVIDIA positions this as an open release, meaning teams can adopt it without vendor lock-in. The combination of an efficient open MoE and a routing layer suggests a shift toward cost-aware agent infrastructure — where the model is no longer a single fixed cost, but a portfolio of options managed by software. Read the full announcement on MarkTechPost.
For buyers and builders, the implication is clear: agent economics are becoming a design decision, not an afterthought. Organizations that adopt routing layers like Switchyard early could gain a cost advantage over competitors still running monolithic models for every step. The open nature of the release also lowers the barrier to experimentation, letting teams benchmark against their own workloads before committing.
