NVIDIA has introduced Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model, alongside NeMo Switchyard, an open-source library for routing requests among multiple AI models. The company says the releases are designed for always-on agentic systems that need to balance accuracy, response time, operating cost and deployment control.

The launch reflects a shift in how NVIDIA is positioning open models. Rather than presenting Lightning as a standalone replacement for every frontier model, the company describes it as a specialist that can handle high-volume tasks inside larger systems. Switchyard is intended to decide which model should handle each step of an agent workflow, potentially reducing the number of requests sent to more expensive models.

A specialist model for multi-agent systems

Nemotron 3.5 Lightning is positioned for targeted workloads such as code review, tool use, security alert monitoring and billing support. NVIDIA says larger reasoning models can plan or coordinate an operation while smaller specialized models execute individual tasks. That architecture is increasingly relevant to developers building agents that run continuously rather than responding to occasional chat prompts.

The model is open and customizable, according to NVIDIA. Organizations can post-train it with NVIDIA NeMo using their own domain data, tools and workflows. NVIDIA also released Nemotron-RL-Agentic-Terminal-Pivot, a reinforcement-learning dataset used in post-training the model for coding-agent capabilities.

The company says Lightning can deliver up to four times faster output and complete agentic tasks 30% faster than other models in its class. Those figures are NVIDIA-reported performance claims, not independently verified results in the supplied evidence. The comparison set and testing conditions are not detailed in the available announcement, so buyers will need to validate the benefits against their own prompts, tools and hardware.

NVIDIA says the model can run across NVIDIA RTX PCs, DGX Spark, DGX Station, Jetson systems, RTX PRO workstations, data centers and cloud environments. That range matters for teams handling sensitive data or seeking to use existing infrastructure rather than sending every request to a hosted API.

Switchyard turns model choice into infrastructure

NeMo Switchyard addresses a practical problem in agent development: different tasks often require different models, but manually managing those decisions can become a large integration burden. NVIDIA’s library can route requests across an organization’s mix of open, proprietary and NVIDIA models without requiring developers to rewrite their applications.

The router can be tuned for priorities including quality, latency and cost. In a system that uses a frontier model for planning but a smaller model for routine execution, this approach could make model selection part of the application’s runtime rather than a fixed architectural decision.

NVIDIA’s framing also points to a broader change in AI infrastructure. As agents make multiple calls during a single workflow, token usage and latency can compound quickly. A routing layer that sends only difficult requests to premium models may be more valuable than a modest improvement in a single model’s benchmark score, particularly for customer support, coding, security and back-office automation.

Switchyard is being made available as an open-source library. NVIDIA says it is working with ecosystem companies to integrate routing into existing developer tools and platforms, including Kong AI Gateway, LiteLLM and Nous Research’s Hermes. LangChain is also cited as an integration partner, while Cognition has tested a staged router in Devin Desktop for NVIDIA internal use.

What the performance evidence shows — and does not show

NVIDIA reports that internal benchmarks kept frontier-level accuracy while reducing task-completion cost to nearly one-third of using Opus 4.8 alone. Because the benchmark is internal and vendor-controlled, the claim should be treated as directional rather than a general guarantee. Results will depend on the model pool, routing policy, workload mix and how much quality degradation an organization accepts.

Partner examples offer additional, but still vendor-supplied, signals. NVIDIA says Boomi achieved 100% domain-routing accuracy across five routing capabilities, sent 59% of traffic to a fine-tuned model that was five times faster, and reduced later-turn latency by 21%. Classmethod reported an initial 27% cost reduction while maintaining quality, while Cognition reported a 28% reduction in mean cost compared with sending all requests to one frontier model.

Other cited results vary. NVIDIA says Cadence improved efficiency by 9.9% in a formal-verification use case. LangChain reported 74% lower cost across 145 multi-turn Deep Agents tasks by sending 7% of calls to a frontier model, with a 6% accuracy tradeoff. These figures illustrate the central tradeoff in routing: savings may be substantial, but lower cost can come with changes in accuracy, latency or workflow behavior.

NVIDIA also names CrowdStrike, Harvey with Trajectory, CodeRabbit with Baseten, Lila Sciences and Fastino Labs as organizations customizing Nemotron 3.5 Lightning. The announcement does not provide independent customer evaluations, production scale, contract details or adoption metrics. The strongest claims about speed, cost and customer performance therefore remain attributable to NVIDIA and its cited partners.

Why this matters to AI builders and enterprises

For builders, the combination of a customizable model and a routing layer could reduce dependence on a single provider. Teams can assign sensitive or repetitive tasks to a locally deployed model while retaining access to proprietary systems for harder cases. That can support privacy requirements, lower recurring inference costs and give engineering teams more control over model updates.

The tradeoff is operational complexity. A router must be evaluated not only on average accuracy but also on failure handling, tool-call reliability, context retention and behavior across multi-turn tasks. A seemingly small accuracy decline can become material when an agent makes dozens of decisions in a workflow. Organizations will also need observability that explains why a request was routed to a particular model and whether that decision improved the outcome.

For enterprise buyers, Switchyard’s value will depend on interoperability and governance. Routing across models from different vendors can reduce lock-in, but it also introduces more components to secure, monitor and audit. NVIDIA’s emphasis on publishing training data and techniques where licensing allows may help researchers and governance teams investigate the model, although disclosure does not by itself establish production reliability.

The release also sharpens competition around the AI application stack. Model providers are increasingly competing not just on raw capability but on the control plane around inference: routing, fine-tuning, deployment location and cost management. NVIDIA is using its hardware ecosystem and NeMo software to connect those layers from local PCs to cloud infrastructure.

What to watch next

The first signal will be independent testing of Nemotron 3.5 Lightning on coding, security and enterprise tool-use workloads, including comparisons with similarly sized open models. Published evaluations should clarify whether NVIDIA’s speed claims hold across different hardware and long-context agent tasks.

Developers should also watch how widely NeMo Switchyard appears in production frameworks, especially whether integrations such as LiteLLM and Kong reduce setup costs for existing applications. Evidence about routing failures, accuracy tradeoffs and monitoring capabilities will be more useful than headline savings alone.

Finally, enterprise adoption will depend on deployment economics. The important questions are whether local and on-premises execution can match hosted alternatives, how much post-training is required for domain accuracy, and whether model routing produces reliable savings after infrastructure, evaluation and governance costs are included.

Creati.ai perspective

NVIDIA’s announcement is most significant as an infrastructure bet: agentic applications may be built from coordinated model specialists rather than one universal model. Nemotron 3.5 Lightning supplies a customizable specialist, while NeMo Switchyard addresses the runtime decision of when to use it.

The opportunity is credible, but the evidence remains largely NVIDIA-reported. For builders, the practical test is not whether routing lowers an isolated benchmark cost; it is whether it preserves end-to-end task reliability while making deployment, privacy and model operations easier to manage.