NVIDIA Launches Nemotron 3.5 Lightning and NeMo Switchyard for Agent Routing
The 30-billion-parameter open model targets specialised agent tasks, while an open routing library directs requests across mixed-model systems.
NVIDIA announced Nemotron 3.5 Lightning and NeMo Switchyard on August 11, expanding its open-model stack for agent systems that divide work across several models. The official release describes a lightweight model for repeated specialised tasks and a separate routing library that selects which model should handle each request.
A smaller model for specialised agent work
Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model. NVIDIA positions it as a component inside larger multi-agent systems rather than as the only model in an application. A more capable reasoning model can plan a workflow, while Lightning handles narrower, higher-volume jobs such as tool use, code review or operational questions.
That architecture matters because agent products do not necessarily need the largest available model for every step. Matching a smaller model to a bounded task can reduce the compute devoted to routine requests while keeping a larger model available for planning or difficult exceptions. The release therefore concerns system design as much as a single model checkpoint.
NVIDIA says Lightning can run across PCs, workstations, data centres and cloud infrastructure. The company also says organisations can customise it with NVIDIA NeMo using their own domain data, tools and workflows. Those claims describe deployment options; they do not establish that one configuration will suit every latency, privacy or accuracy requirement.
NeMo Switchyard separates routing from application code
NeMo Switchyard is an open-source routing library intended for popular agent tools. According to NVIDIA, developers can define a router for their own requirements and direct each request to the model judged most suitable without rewriting the application around every model change.
This separation could make mixed-model systems easier to operate. A team can revise routing policy or introduce a specialised model while preserving the interface used by the rest of the agent. The practical result will still depend on the quality of the router's evaluation criteria, the available models and the cost of mistakes when a request is sent to an unsuitable model.
NVIDIA presents Lightning and Switchyard as complementary releases: one supplies a compact specialist, while the other supplies the control layer that places work across a model ensemble. Neither component removes the need to test the full agent workflow, including tool permissions, failure recovery and the quality of final outputs.
Performance claims need workload-specific testing
NVIDIA reports up to four-times faster output and 30% faster agentic task completion than other models in Lightning's class. Those figures come from NVIDIA's release and its cited evaluations; AI News Bank did not independently reproduce them. They should be read as vendor results rather than a universal speed guarantee.
The useful comparison for adopters will be end-to-end performance on their own tasks. Token generation speed can help, but an agent may still lose time through tool calls, retrieval, retries or poor routing. Accuracy also matters: a faster specialist that requires more corrections may not improve the overall workflow.
Status
Confirmed product release from NVIDIA. Internal confidence is high because the model, library, intended roles and availability are documented by the primary actor. Performance comparisons remain attributed vendor claims.
Sources
Update note: Last reviewed 2026-08-11. We will revise this article if NVIDIA publishes additional model documentation, independent evaluations or material availability changes.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.