NVIDIA adds Nemotron 3.5 Lightning and NeMo Switchyard to its open-model lineup

NVIDIA has expanded its Nemotron open-model family with Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model aimed at high-volume agentic AI workloads, alongside the release of NeMo Switchyard, an open-source model routing library. The two pieces address a structural problem in agent design: as AI systems move from single chatbots to multi-model architectures, deciding which model handles which request becomes its own engineering task.

Lightning targets the "systems of models" pattern, where a larger reasoning model plans and orchestrates a workflow while smaller specialized models handle tasks like code review, tool use, and security alert monitoring. The model is open and can be post-trained with NVIDIA NeMo on an organization's own domain data. NVIDIA also released Nemotron-RL-Agentic-Terminal-Pivot, an agentic reinforcement learning dataset used to train the model for coding agent capabilities.

NVIDIA reports up to 4x faster output speed and 30% faster agentic task completion compared with other models in its class, citing its PinchBench benchmarks. Those figures are NVIDIA's own measurements rather than independent evaluations. Deployment spans RTX PCs, DGX Spark and Station systems, edge devices, data centers, and cloud environments, which supports organizations that need local execution for latency or privacy reasons. As with prior Nemotron releases, NVIDIA publishes training data and techniques where licensing permits.

NeMo Switchyard sits between an agent application and the models it calls, directing each request to the most suitable model based on quality, latency, or cost priorities, without requiring developers to rewrite their applications. The router works across a mix of open, proprietary, and NVIDIA models. NVIDIA's internal benchmark claims Switchyard maintains frontier-level accuracy while reducing task completion cost to nearly one-third of routing everything to a single frontier model.

The rationale is that a single default model either overspends on easy tasks or underperforms on hard ones, and manual routing becomes integration work that slows deployment. Several partners report early results, all self-reported:

• Boomi reports 100% domain-routing accuracy and sending 59% of traffic to a 5x faster fine-tuned model, reducing later-turn latency by 21%.
• LangChain reports 74% lower cost across 145 multi-turn agent tasks by routing only 7% of calls to a frontier model, at a 6% accuracy tradeoff.
• Ramp reports matching frontier model performance in its SWE-Bench evaluation while cutting costs by 58% and runtime by 33%.
• Cognition integrated the router into Devin Desktop and reports a 28% mean cost reduction relative to a single frontier model.

The LangChain result is notable: routing produced meaningful cost savings, but not without an accuracy cost. How that tradeoff scales beyond 145 tasks remains to be seen.

Nemotron 3.5 Lightning is available through Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as a NIM microservice, plus partner platforms. NeMo Switchyard is on GitHub, with integrations from Kong, LiteLLM, LangChain, and Nous Research either live or in progress.