Nvidia has expanded its Nemotron model family with Nemotron 3.5 Lightning and released the NeMo Switchyard routing library, targeting AI systems that use multiple models for autonomous agent tasks.
Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts open model designed for specialised work inside larger agent systems. It is aimed at long-running, high-volume tasks such as code review, tool use, security alert monitoring and billing support.
Nvidia also introduced NeMo Switchyard, an open-source library for routing models inside agent tools. It is designed to direct requests to different models based on the task, including open, proprietary and Nvidia models, without requiring developers to rewrite applications.
Model strategy
The announcements reflect a broader shift in AI development toward systems that combine several models rather than relying on a single general-purpose model. In this setup, a larger reasoning model can plan or orchestrate a workflow, while smaller models handle narrower tasks that require lower cost or faster responses.
Nvidia positions its Nemotron open models around that architecture. Nemotron 3.5 Lightning follows the earlier Nemotron 3 Nano release and is intended to sit alongside larger reasoning models such as Nemotron 3 Ultra.
According to Nvidia, Nemotron 3.5 Lightning delivers up to four times faster output speed than other models in its class, leading to 30% faster agentic task completion. Nvidia also cited PinchBench results showing frontier-level accuracy compared with rival models in the same category.
The model is fully customisable and can be post-trained with Nvidia NeMo using an organisation's own data, tools and workflows. That allows companies to adapt it for domain-specific work where accuracy matters more than broad general knowledge.
Customer use
Nvidia cited a range of companies customising the model for specific workloads, including CrowdStrike for cybersecurity, Harvey with Trajectory for legal services, and CodeRabbit with Baseten for code review.
Lila Sciences is using the model to improve reasoning for tasks in the physical and life sciences. Fastino Labs has also customised it for software development, finance and healthcare workloads, where users are reporting strong accuracy on specialised tasks, according to Nvidia.
The model can run on local systems including RTX PCs, DGX Spark, DGX Station and Jetson devices, as well as across workstations, edge systems, data centres and cloud environments. That gives organisations a way to keep some AI work on premises where privacy, latency or infrastructure constraints matter.
Nvidia also released Nemotron-RL-Agentic-Terminal-Pivot, a reinforcement learning dataset used to post-train the model for coding agent work. It said it publishes as much training data and training technique detail as licensing permits to improve traceability and auditing.
Routing economics
NeMo Switchyard addresses a separate issue for companies building AI agents: choosing which model should handle each step in a workflow. Different models can be better at coding, reasoning, lightweight requests or local deployment, and using one default model for every task can raise costs or reduce quality.
Nvidia said developers can tune or modify Switchyard with different routing algorithms based on quality, latency and cost requirements. Internal benchmarks cited by the company said the library maintained frontier-level accuracy while reducing task completion cost to nearly one-third of using Opus 4.8 alone.
Several software groups have already tested or integrated the routing library. Boomi evaluated Switchyard across five routing functions and, according to Nvidia, achieved 100% domain-routing accuracy, sent 59% of traffic to a fine-tuned model that was five times faster, and reduced later-turn latency by 21%.
Cadence said it improved efficiency by 9.9% in a formal verification use case using its ChipStack AI Super Agent. Classmethod is running opencode and Fireworks workloads with Switchyard internally and reported an initial 27% cost reduction while maintaining quality.
Cognition integrated the staged router into Devin Desktop for Nvidia internal use and achieved near-frontier performance on FrontierCode Main while reducing mean cost by 28% compared with routing all requests to a single frontier model. Kong is delivering routing with NeMo Switchyard through Kong AI Gateway, while LiteLLM is adding it as a plug-in to its proxy layer.
LangChain reported 74% lower cost across 145 multi-turn Deep Agents tasks by routing only 7% of calls to a frontier model, with a 6% accuracy trade-off. Ramp used the software to match a frontier model's performance while cutting costs by 58% and runtime by 33% in Ramp SWE-Bench.
Nous Research has integrated NeMo Switchyard into Hermes, and Siemens is benchmarking the software for its Fuse EDA AI Agent. These examples point to a growing market for routing tools that manage trade-offs between cost, speed and output quality across increasingly complex AI stacks.
Nemotron 3.5 Lightning is being distributed through model repositories and Nvidia's own services, while NeMo Switchyard is available as open-source software.