NVIDIA has expanded its Nemotron family with Nemotron 3.5 Lightning, a lightweight open model designed to handle specialized tasks inside larger AI-agent systems.
The model uses a 30-billion-parameter mixture-of-experts architecture and is aimed at high-volume workloads such as code review, tool use, security-alert monitoring and other repeatable tasks that may not require a large frontier model for every step.
NVIDIA is launching the model alongside NeMo Switchyard, an open-source routing library for agent applications. Switchyard can direct individual requests to different models according to requirements such as capability, latency and cost, including combinations of open, proprietary and NVIDIA models.
The underlying idea is that an advanced agent does not need to be a single model. A larger reasoning model can plan or orchestrate a workflow, while smaller models perform narrower jobs repeatedly. Routing software then decides which model should handle each step.
NVIDIA says Nemotron 3.5 Lightning is designed to make that architecture more efficient. The company reports faster output and shorter task-completion times compared with other models in its class, although those performance figures come from NVIDIA's own evaluations and will need broader independent testing.
The open-model approach also gives organizations more control over deployment. NVIDIA says Lightning can be customized with domain data and run across local systems, workstations, data centers and cloud infrastructure, allowing companies to keep specialized workloads closer to their own environments when needed.
For enterprises, the more consequential release may ultimately be the router rather than any single model. If agents increasingly rely on a collection of models, software that can decide when an expensive frontier model is necessary — and when a smaller model is sufficient — becomes part of the core infrastructure of the application.
