Nvidia releases Nemotron 3.5 Lightning and NeMo Switchyard to give enterprise AI capability options

Nvidia has rolled out two more pieces for enterprise AI teams that are trying to make agents less like a science project and more like a tool people can actually use. One is Nemotron 3.5 Lightning, a customizable open model built for high-volume work. The other is NeMo Switchyard, an open-source router that helps pick which model should handle each step of an agent’s workflow.

Lightning sits in Nvidia’s open Nemotron family and is aimed at always-on agents that need to churn through lots of simpler tasks. Nvidia says the model has 30 billion parameters and is a mixture-of-experts design. The company claims it can deliver up to four times the output speed and 30% faster agentic task completion than other models in its weight class.

The bigger point is that Nvidia is pushing the idea that enterprise AI is no longer just a contest over raw capability. A fast, cheaper model can be the right answer for one step, while a heavier reasoning model fits another. That is exactly the sort of messy, mixed workload companies are starting to see as agents spread.

Nvidia says Lightning is easy to specialize. Briski said CodeRabbit used Nvidia’s standard auto model recipe, trained for one epoch, and built a router agent for $85 in about two hours. She also described another partner dropping Lightning into an existing post-training stack without changes, plus one case where a team trained on a single H100 card, and another where the job ran overnight and was ready in the morning.

Nvidia is also releasing its datasets and the post-training recipes and frameworks used to build the models. That gives enterprises a way to combine Nvidia’s data with their own proprietary material, tools and workflows. NeMo Switchyard then sits on top as the traffic cop, routing prompts based on things like quality, delay and cost, and Nvidia says it is already working with partners including Boomi, Cadence Design Systems, Classmethod, Cognition AI, Kong, Langchain, Nous Research and Siemens.