TECH NEWS | NVIDIA expands Nemotron lineup with agentic AI model, routing tool
Nemotron 3.5 Lightning is intended to operate as part of multi-model AI systems, where larger reasoning models handle planning and orchestration.
NVIDIA has expanded its open AI model lineup with Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model designed for specialized tasks in long-running agentic AI workflows.
The company announced the model Aug. 11 alongside NeMo Switchyard, an open source model-routing library designed to direct requests to different AI models based on factors such as quality, latency and cost.
Nemotron 3.5 Lightning is intended to operate as part of multi-model AI systems, where larger reasoning models handle planning and orchestration while smaller specialized models perform targeted tasks such as code review, tool use, security monitoring and billing-related queries.
NVIDIA said Nemotron 3.5 Lightning can deliver up to four times faster output and enable 30% faster agentic task completion compared with other models in its class. The company cited PinchBench results showing faster task completion while maintaining what it described as frontier-level accuracy.
The model is open and customizable, allowing organizations to post-train it using their own domain data, tools and workflows through NVIDIA NeMo. NVIDIA said the model can be deployed locally on systems including RTX PCs, DGX Spark, DGX Station and Jetson, as well as on RTX PRO workstations, data centers and cloud environments.
Companies including CrowdStrike, Harvey, CodeRabbit, Lila Sciences and Fastino Labs are using or customizing Nemotron 3.5 Lightning for applications including cybersecurity, legal services, code review, science, finance and health care, according to NVIDIA.
NVIDIA is also releasing Nemotron-RL-Agentic-Terminal-Pivot, a reinforcement learning dataset used to post-train the model for coding-agent capabilities.
NeMo Switchyard addresses another challenge in agentic AI: choosing which model should handle a particular task. NVIDIA said the open source library can automatically route requests among open, proprietary and NVIDIA models without requiring developers to rewrite their existing applications.
Developers can modify routing algorithms according to priorities such as accuracy, latency and cost. NVIDIA’s internal benchmarks showed Switchyard maintaining frontier-level accuracy while reducing task-completion costs to nearly one-third of using Opus 4.8 alone.
NVIDIA said partners including Boomi, Classmethod, Cognition, LangChain and Ramp have tested or integrated Switchyard, reporting reductions in cost or latency in various workloads.
Nemotron 3.5 Lightning is available through Hugging Face, ModelScope, OpenRouter and NVIDIA’s build.nvidia.com platform, as well as through NVIDIA cloud and ecosystem partners. NeMo Switchyard is available on GitHub.
