NVIDIA has released NeMo Switchyard, an open source routing library that decides which model handles each step of an agent workflow, and paired it with a new 30-billion-parameter open model called Nemotron 3.5 Lightning. Both were announced in NVIDIA’s Nemotron 3.5 Lightning and NeMo Switchyard post.
The routing piece is the more structural change. Teams building agents usually pick one default model and live with the bill, or hand-wire routing rules that rot as models change. According to NVIDIA, Switchyard sits inside existing agent tools and directs each prompt to the most suitable model across a developer’s own mix of open, proprietary and NVIDIA models without rewriting the application, and the routing algorithm can be tuned for quality, latency or cost.
What the partner numbers say
The efficiency figures come from vendors, not from independent testing. NVIDIA’s internal benchmarks claim Switchyard holds frontier-level accuracy while cutting task completion cost to nearly one-third of Opus 4.8 alone, and NVIDIA reports partner results including LangChain measuring 74% lower cost across 145 multi-turn Deep Agents tasks by sending only 7% of calls to a frontier model at a 6% accuracy tradeoff, Ramp cutting costs 58% and runtime 33% on Ramp SWE-Bench, and Cognition reducing mean cost 28% in Devin Desktop, all per NVIDIA’s announcement. Boomi, Cadence, Classmethod, Kong, LiteLLM, Nous Research and Siemens are also listed as integrating or evaluating the router in that same post.
The model beside the router
Nemotron 3.5 Lightning is a mixture-of-experts model, an architecture that activates only a subset of its parameters per token, as NVIDIA’s own glossary explains. NVIDIA states the model delivers up to 4x faster output speed and 30% faster agentic task completion versus others in its class, and that it runs locally on RTX PCs, DGX Spark, DGX Station and Jetson as well as in data centers and cloud, per NVIDIA. NVIDIA also published Nemotron-RL-Agentic-Terminal-Pivot, a reinforcement learning dataset used to post-train the model for coding agents. The company’s accompanying explainer video, Why AI Agents Need More Than One Model, frames the same ensemble argument.
Why it matters
The pitch is that a small specialized model plus a router beats one large default model on cost without collapsing quality. That is a claim worth testing on your own traffic before trusting the partner percentages, because routing gains depend entirely on how much of your workload is genuinely easy. Nemotron 3.5 Lightning is available on Hugging Face, ModelScope, OpenRouter and build.nvidia.com as an NIM microservice, and Switchyard is on GitHub, according to NVIDIA. For a distribution angle on the same model, see our report on how Ubuntu adds NVIDIA Nemotron 3.5 Lightning in one command.
