The 30-billion-parameter open model is designed for high-volume agentic AI tasks, while NVIDIA’s NeMo Switchyard library automatically routes requests between models based on factors such as quality, latency and cost.
NVIDIA has expanded its Nemotron 3 model family with Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model designed for specialised tasks within larger multi-agent systems. The model is built for high-volume workloads such as code review, tool use, security alert monitoring and answering billing questions.
Nemotron 3.5 Lightning can be customised using NVIDIA NeMo and an organisation’s data, tools and workflows. According to Nvidia, the model delivers up to four times faster output speed and enables 30 per cent faster agentic task completion as compared to similar frameworks available. The model can run locally or on-premises depending upon the user requirements and works well with various platforms including NVIDIA RTX PC, DGX Spark, DGX Stations, and Jetsons, edge infrastructure, RTX PRO workstations, data centers, and cloud environments.
NVIDIA has also released NeMo Switchyard, an open-source model-routing library for AI agents. NeMo Switchyard intelligently routes prompts to appropriate models at each stage of an agent workflow. Switchyard is flexible to configure as well and can adopt various routing strategies as per need based on parameters such as quality, latency, and cost. According to NVIDIA’s Internal benchmarks, the Switchyard maintains frontier-level accuracy while reducing task completion cost to nearly one-third of using Opus 4.8 alone.
Training information and techniques are provided to the extent permitted by licensing. Along with the release of Lightning, NVIDIA is introducing the Nemotron-RL-Agentic-Terminal-Pivot dataset, which was used to post-train coding-agent capabilities. The model was developed with contributions from the Nemotron Coalition including evaluation methodologies, inference software and datasets.
The Nemotron 3.5 Lightning is available through platforms including Hugging Face, ModelScope, OpenRouter, NVIDIA’s build.nvidia.com platform as an NVIDIA Nim microservice along with NVIDIA Cloud partners and other platforms. NeMo Switchyard is available on GitHub.
