Nvidia Ships 30B Nemotron Lightning, Opens Switchyard Router

Nvidia shipped Nemotron 3.5 Lightning today, a 30-billion-parameter mixture-of-experts model built specifically for the agent-heavy workloads its own customers keep asking about. Nvidia's blog post frames Lightning as a specialized component inside larger multi-agent systems, not a general-purpose chat model. The pitch is that if you are running long autonomous agents at volume, a fast small MoE routed intelligently beats a single expensive frontier call.

The claimed numbers are about speed rather than raw intelligence. Nvidia says the model delivers up to 4x faster output and about 30% faster agentic task completion against unnamed models "in its class," with what it calls frontier-level accuracy on its own PinchBench suite. Alongside the weights it open-sourced NeMo Switchyard, a routing library that shuttles agent traffic across mixed models, pitched as bringing task completion cost "to nearly one-third of Opus 4.8 alone."

The customer roster reads like a checklist of who is already spending serious money on agents. Cognition wired Switchyard into Devin Desktop for Nvidia internal use and reports a 28% mean cost reduction versus routing every request to a single frontier model. Ramp saw 58% cost and 33% runtime cuts, LangChain traded 6% accuracy for 74% lower cost, and Cadence, CrowdStrike, Harvey with Trajectory, CodeRabbit with Baseten, Fastino Labs and Lila Sciences all appear as design partners. This lands amid a run of open-weight releases our tracker has been logging across recent open-source AI news. Motif detailed its 314B MoE with 13.2B active per token the same day.

The comparison numbers are all Nvidia's, on benchmarks Nvidia named, and the post is careful with what it actually publishes, releasing "as much of the training data and techniques as licensing permits." It never states the active-parameter count for the MoE, never names the specific open-weights license that attaches, and never benchmarks Lightning against particular rival open MoE models the way independent evaluators would. Cognition's 28% is the only externally-attributed cost figure in the release.

What makes this interesting is that Switchyard is open source and Lightning runs locally on RTX PCs, DGX Spark, DGX Station and Jetson, meaning the whole agent stack can move off hosted APIs for teams willing to run their own hardware. Whether Anthropic and OpenAI respond by cutting their agent-tier pricing is what decides how much of that savings survives the quarter.