Nvidia Releases Nemotron 3.5 Lightning to Cut AI Agent Costs

Nvidia has released Nemotron 3.5 Lightning, an open 30-billion-parameter mixture-of-experts model built to run the fast, repetitive steps inside multi-step AI agents, alongside NeMo Switchyard, an open-source router that sends each task to whichever model handles it best. The company announced both releases on August 11 via its engineering blog.

Built for the busy end of agent work

According to Nvidia, Nemotron 3.5 Lightning is meant for the high-volume “execution layer” of agentic systems — the many small subtasks, like tool calls or document lookups, that a larger reasoning model would otherwise handle at higher cost. The company says the model delivers up to four times the output speed of similarly sized models and completes agentic tasks about 30% faster than comparable models, while holding “frontier-level accuracy” on its internal PinchBench benchmark suite.

The release joins a fast-growing field of open-weight models sized for agent execution rather than general chat — Meta’s 30-billion-parameter Muse Glimmer, released weeks earlier, targets similar ground.

A router, not just a model

NeMo Switchyard is arguably the bigger idea: it’s a routing layer developers drop into existing agent code to automatically split work across a mix of open, proprietary, and Nvidia models by quality, latency, or cost needs, without rewriting the application. Nvidia says internal testing found the combination held frontier-level accuracy while cutting the cost of completing tasks to roughly a third of using Anthropic’s Opus 4.8 model alone.

Code-review startup CodeRabbit used Nvidia’s standard training recipe to fine-tune a router agent for its own workflow in about two hours, at a cost of roughly $85 for one training epoch, according to Nvidia.

Both releases are open-source: Nemotron 3.5 Lightning is available on Hugging Face, ModelScope, OpenRouter, and Nvidia’s own build.nvidia.com platform, and can run locally on Nvidia’s RTX PCs, DGX Spark, DGX Station, and Jetson hardware. NeMo Switchyard’s code is on GitHub, with partner-platform integrations expected to follow.