Nvidia releases Nemotron 3.5 Lightning and NeMo Switchyard
Nvidia’s new open model and routing library target enterprise AI agents that need to balance specialized work, quality, latency and cost.
By Dominic Okoye · Staff Writer
· 3 min read
Nvidia has released Nemotron 3.5 Lightning, a 30-billion-parameter model for high-volume agent tasks, alongside NeMo Switchyard, an open-source library intended to select models for each step of an AI-agent workflow. The Nvidia Nemotron Lightning Switchyard release pairs a specialized execution model with a routing layer, reflecting a practical enterprise problem: a single default model can be unnecessarily expensive for routine work and insufficient for harder reasoning tasks.
Nvidia did not disclose pricing, availability terms or independent benchmark results in the material reviewed. Its performance and cost figures are company claims.
What do Nvidia Nemotron 3.5 Lightning and NeMo Switchyard do?
Lightning is Nvidia’s addition to its Nemotron 3 family for repeatable, high-volume work inside always-on, multi-agent systems. Nvidia describes it as an open, customizable mixture-of-experts model, an architecture that activates only a subset of its components for a given request. The company positions it for tasks such as tool use, code review, security-alert monitoring and billing questions, rather than the planning and orchestration work assigned to larger reasoning models.
Nvidia says organizations can post-train Lightning with its NeMo software using their own domain data, tools and workflows. The company also says it is releasing the Nemotron-RL-Agentic-Terminal-Pivot dataset, used in Lightning’s post-training for coding-agent capabilities, and will publish training data and techniques where licensing permits.
Deployment is part of the pitch. Nvidia says Lightning can run locally on RTX PCs, DGX Spark, DGX Station and Jetson systems, and can extend to edge devices, RTX PRO workstations, data centers and cloud environments. Those options could matter to companies that want to keep specialized workloads on infrastructure they control, but Nvidia did not provide comparative operating-cost data for those configurations.
Routing is the control layer
NeMo Switchyard is separate from Lightning. Nvidia describes it as a library that sends prompts or individual workflow steps to a model selected for the task, without requiring developers to rewrite their agent applications. Developers can alter its routing algorithms around priorities including answer quality, latency and cost.
In practice, that means a company could use Lightning for frequent, bounded work while reserving a more capable model for complex planning or reasoning. Nvidia previously described Nemotron 3 Ultra as a 550-billion-parameter mixture-of-experts model, with 55 billion active parameters, built for long-running agent orchestration. The two releases therefore form a system-of-models approach: one layer executes specialized work and another decides where work should go.
Nvidia claims Lightning can produce output up to four times faster and complete agentic tasks 30% faster than other models in its class. It also says internal Switchyard tests preserved “frontier-level” accuracy while cutting task-completion cost to nearly one-third of using Opus 4.8 alone. Nvidia has not supplied independent validation or methodology in the available announcement material, so enterprises will need to test those claims against their own prompts, data and serving costs.
- Nemotron 3.5 Lightning: a customizable model for high-volume, specialized agent tasks.
- NeMo Switchyard: a routing library for choosing among models according to task and operational constraints.
- Enterprise implication: Nvidia is offering components for mixed-model agent stacks rather than arguing that one model should handle every request.
This story draws on original reporting from SiliconANGLE.
