NVIDIA Releases Nemotron 3.5 Lightning and NeMo Switchyard for Agentic AI

NVIDIA has released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model, alongside NeMo Switchyard, an open-source model routing library for agentic AI applications.

The two releases target systems where multiple AI models handle different parts of a workflow. Nemotron 3.5 Lightning is intended for specialised, high-volume tasks, while Switchyard routes requests to different models based on factors such as quality, latency and cost.

NVIDIA said Nemotron 3.5 Lightning can deliver up to four times faster output and 30% faster agentic task completion than other models in its class.

The model is available through Hugging Face, ModelScope, OpenRouter and NVIDIA's build.nvidia.com platform. NeMo Switchyard is available on GitHub.

The company said the model is built for “high-volume tasks powering always-on agents” and can be customised using NVIDIA NeMo with an organisation’s own data, tools and workflows.

NVIDIA sees the model working alongside larger reasoning models rather than replacing them. “A frontier reasoning model such as Nemotron 3 Ultra or GPT-5.6 may plan and orchestrate a workflow, while smaller specialised models like Nemotron 3.5 Lightning can perform targeted tasks such as code review, tool use, security alert monitoring and answering billing questions,” the company said.

The model is being customised for use cases including cybersecurity, legal services, code review, science, finance and healthcare. NVIDIA cited work by CrowdStrike, Harvey with Trajectory, CodeRabbit with Baseten, Lila Sciences and Fastino Labs.

Nemotron 3.5 Lightning can also run locally on NVIDIA RTX PCs, DGX Spark, DGX Station and Jetson devices, as well as on RTX PRO workstations, data centres and cloud infrastructure.

NVIDIA is also releasing the Nemotron-RL-Agentic-Terminal-Pivot dataset, which was used to post-train the model for coding-agent capabilities.

NVIDIA targets model costs with Switchyard

NeMo Switchyard addresses another problem in multi-model AI systems: deciding which model should handle each request.

Instead of sending every task to a single model, Switchyard can route requests to different models according to an application's requirements. Developers can modify the routing algorithms to prioritise quality, latency or cost.

“Some models are better for coding, some for reasoning, some for lightweight tasks and some are optimised to run locally for greater privacy and efficiency,” NVIDIA said.

The company said internal benchmarks showed Switchyard maintaining frontier-level accuracy while reducing task-completion costs to nearly one-third of using Opus 4.8 alone.

Early partner tests have produced further cost and latency reductions. LangChain reported a 74% reduction in cost across 145 multi-turn Deep Agents tasks by sending only 7% of calls to a frontier model, with a 6% accuracy trade-off. 

Cognition integrated Switchyard into Devin Desktop for NVIDIA’s internal use and reported a 28% reduction in mean cost compared with routing all requests to a single frontier model.

Switchyard is also being integrated with tools and platforms including Kong AI Gateway, LiteLLM, Hermes and Devin Desktop.