NVIDIA announced Nemotron 3.5 Lightning on August 11, 2026, alongside NeMo Switchyard, an open-source library for routing requests between models used by AI agents. The model targets long-running agentic workloads; the router is designed to decide which model should handle each request.
A model chosen for each task
Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model. NVIDIA says it delivers up to four times the output speed of other models in its category and reduces agentic task completion time by 30%. Those figures come from NVIDIA’s own comparisons, not an independent benchmark, so they indicate the company’s expected advantage rather than a universal result.
NeMo Switchyard adds a separate routing layer. It can send each request to the most suitable model in a pool containing open, proprietary and NVIDIA models. Developers can tune the routing logic around quality, latency and cost, and NVIDIA says existing applications do not need to be rewritten to use the library.
Cybersecurity is one of the named use cases
The announcement has a direct cybersecurity angle because NVIDIA lists CrowdStrike among the organizations customizing Nemotron 3.5 Lightning. Other examples include Harvey with Trajectory for legal services and CodeRabbit with Baseten for code review. CrowdStrike is the clearest indication that NVIDIA is presenting the model as infrastructure for specialized organizational workflows, rather than only as a general-purpose interface.
NVIDIA says organizations can post-train Nemotron 3.5 Lightning with NeMo using their own data, tools and workflows. That could help teams adapt the model to internal operating patterns. The announcement does not describe CrowdStrike’s resulting system, its security controls or the tasks it has deployed, however, so the cybersecurity example does not yet establish what the model can do in a real production environment.
The reported routing gains vary by deployment
NVIDIA’s internal benchmarks say NeMo Switchyard can maintain frontier-level accuracy while reducing a task’s execution cost to nearly one-third of using Opus 4.8 alone. The partner results cited in the announcement show why the outcome depends on the workload and routing policy.
Boomi reported 100% domain-routing accuracy across five routing capabilities. It sent 59% of its traffic to a fine-tuned model that was five times faster and reduced later-turn latency by 21%. LangChain reported a 74% cost reduction across 145 multi-turn Deep Agents tasks by sending only 7% of calls to a frontier model, with a 6% accuracy trade-off.
Ramp reported 58% lower costs and 33% faster execution on Ramp SWE-Bench while matching the performance of a frontier model. These figures are attributed to the organizations cited by NVIDIA; they should not be treated as expected results for every deployment. Routing can lower cost and latency, but the trade-off depends on which requests are sent to smaller or fine-tuned models and how quality is measured.
Where developers can use the releases
Nemotron 3.5 Lightning is available through Hugging Face, ModelScope, OpenRouter and build.nvidia.com as an NVIDIA NIM microservice. NVIDIA says it can run locally on NVIDIA RTX PCs, DGX Spark, DGX Station and Jetson, as well as on edge devices, NVIDIA RTX PRO workstations, data centers and cloud environments.
NeMo Switchyard is available on GitHub, with support for partner platforms expected soon. NVIDIA has also published Nemotron-RL-Agentic-Terminal-Pivot, an agentic reinforcement-learning dataset used to post-train coding-agent capabilities.
For cybersecurity organizations, the immediate change is access to an open model and a routing layer that can be adapted to internal data, tools and workflows. The practical decision will depend on whether a team can validate the routing policy, model quality and safeguards against its own tasks; the announcement does not provide that evaluation scope or detailed implementation.
NVIDIA is positioning Nemotron 3.5 Lightning and NeMo Switchyard as infrastructure for selecting a model for each agent task. The reported speed and cost gains are significant signals, but they remain tied to the workloads, routing choices and benchmarks behind each result.