NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard Deliver Faster, Smarter, More Efficient Agentic AI

NVIDIA is pushing harder into the agent era with two releases aimed at the plumbing, not the chatbot demo. Nemotron 3.5 Lightning is the newest member of its Nemotron 3 family, and NVIDIA says it is the highest-efficiency model in its class for long-running agentic workloads. The other piece, NeMo Switchyard, is an open source routing library that can steer requests to the best model inside an agent workflow.

The idea behind Lightning is simple: not every part of an AI system needs a giant reasoning model. NVIDIA frames modern agents as systems of models, where a larger frontier model like Nemotron 3 Ultra or GPT-5.6 can plan the work, while smaller specialists handle things like code review, tool use, security alerts or billing questions. Lightning is a 30-billion-parameter mixture-of-experts model built for those narrower jobs, and NVIDIA says it can produce output up to 4x faster, with agentic task completion 30% faster than other models in its class.

Because it is open and customizable, the model can be post-trained with NVIDIA NeMo on an organization’s own data, tools and workflows. NVIDIA says that should help with accuracy on specialized tasks, and it points to customers and partners already adapting it for cybersecurity, legal services, code review, software development, finance, healthcare and even physical and life sciences. The company also says Lightning can run locally or on premises, including on NVIDIA RTX PCs, DGX Spark, DGX Station and Jetson, or scale out across workstations, data centers and cloud environments.

NeMo Switchyard is meant to make the whole setup less clumsy. Instead of routing everything through one default model or manually stitching together model choices, it can direct each request automatically based on quality, latency and cost targets. NVIDIA’s internal benchmark says that approach keeps frontier-level accuracy while cutting task completion cost to nearly one-third of Opus 4.8 alone. Partners are already claiming real-world gains too, from Boomi’s 100% domain-routing accuracy to Ramp’s 58% cost cut and 33% runtime drop in Ramp SWE-Bench.

Both tools fit NVIDIA’s familiar playbook: give enterprises control, then wrap that control in enough infrastructure that it feels practical. Lightning is available through Hugging Face, ModelScope, OpenRouter and build.nvidia.com, plus NVIDIA NIM and partner platforms. Switchyard is on GitHub now, with partner platforms promised later.