NVIDIA released Nemotron 3.5 Lightning on August 11, a 30-billion-parameter model built for the high-volume execution layer inside always-on AI agents: making tool calls, checking tool outputs, handing subtasks to helper agents, the routine steps where a frontier model is overkill. It is a mixture-of-experts design, so only 3 billion parameters run per token, which keeps it fast and cheap. NVIDIA reports 86% accuracy on PinchBench, an agent-efficiency benchmark, and says it finishes 10,000 tasks 30% faster than Qwen3.6 35B at similar accuracy, with up to 4x the output speed of comparable models. It runs from a desktop DGX Spark to a data center, and the weights are open under the OpenMDW-1.1 license.
The companion piece is NeMo Switchyard, an open-source router that sends each step of an agent’s work to the model best suited to it, usually a cheap one. LangChain benchmarked it over 145 multi-turn agent tasks, routing between Nemotron 3.5 Lightning and Claude Opus 4.8. The small model handled 93% of all calls for 10.4% of the spend, and the total bill fell 74% against running Opus alone, from $11.45 to $3.00 per run, for a roughly six-point accuracy drop. For an always-on support agent, that turns a frontier-model bill into pocket change.
Most of what an agent does is mechanical, and routing exists so that work never touches a frontier model. Agent bills are now set by routing and small models, not the flagship in the news.
Read More: NVIDIA Nemotron 3 Nano Omni: One Open Model for Vision, Audio, and Text at 9x the Throughput
Sources:
- NVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agents (NVIDIA Technical Blog)
- Route AI Agent Workloads Across Models with NVIDIA NeMo Switchyard (NVIDIA Technical Blog)
- How many of your agent’s calls actually need a frontier model? (LangChain)
- NVIDIA Nemotron 3.5 Lightning model card (Hugging Face)
- NeMo Switchyard (GitHub)
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {NVIDIA {Nemotron} 3.5 {Lightning} {Does} the {Grunt} {Work}
of {AI} {Agents,} 74\% {Cheaper} {When} {Routed}},
date = {2026-08-24},
url = {https://toknow.ai/posts/nvidia-nemotron-3-5-lightning-nemo-switchyard-agent-execution/},
langid = {en-GB}
}