NVIDIA open-sources NeMo Switchyard, a router built to keep work away from the frontier model
Alongside Nemotron 3.5 Lightning — a 30B mixture-of-experts model with 3B active parameters and a 1M-token context — NVIDIA released Switchyard, a tuning-free routing library that sends each step of an agent workflow to the cheapest model that can complete it. NVIDIA claims task cost near a third of running Opus 4.8 alone.
The router is the more interesting release. A model is a product; a router is an argument about how the whole stack should be assembled, and this one argues that the frontier model is a planner you should invoke sparingly.
Switchyard ships three tuning-free strategies: an LLM classifier with session affinity, a stage router that reads recent tool activity to infer what phase the workflow is in, and an escalation router that starts cheap and promotes only on sustained difficulty. None require training a routing model, which is the usual reason routing projects die in a drawer.
Lightning is the piece the router points at. Thirty billion total parameters, three billion active, a hybrid Mamba-2 plus MoE plus attention architecture, a million tokens of context, distilled from Nemotron 3 Ultra and tuned for the execution layer — tool calls, retrieval, validation, formatting, classification, summarisation. NVIDIA claims 4x faster output than similar-sized models and 30% faster agentic task completion. It is on Hugging Face, ModelScope, OpenRouter's free tier and build.nvidia.com.
Cursor's own agent swarm data points the same way: cheaper models handle most coding when a frontier model plans the work. Two independent parties converging on the same architecture in the same week is usually how a default gets set.
MarkTechPost — NVIDIA AI Releases Nemotron 3.5 Lightning and NeMo Switchyard → · NVIDIA Developer — Nemotron 3.5 Lightning delivers fast, accurate specialized task execution for long-running agents → · DataNorth — NVIDIA releases Nemotron 3.5 Lightning and NeMo Switchyard →