NVIDIA releases Nemotron 3.5 Lightning and an open-source router to cut multi-agent costs

NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model with 3 billion active parameters, on August 11, 2026, built to handle specialized tasks inside larger multi-agent systems rather than a standalone frontier model.

The company said the model generates tokens up to four times faster than other models in its class, completing agentic tasks about 30% faster at comparable accuracy. It shipped alongside NeMo Switchyard, an open-source routing library that directs each agent request to the most suitable model on cost, latency or quality.

The pairing targets the economics of agentic AI, where sending every step to a frontier model is expensive. NVIDIA said an escalation-router setup cut cost to nearly a third of using Anthropic’s Claude Opus 4.8 alone while maintaining frontier-level accuracy, citing a 74% cost reduction across 145 multi-turn agent tasks in testing with the LangChain framework by routing only 7% of calls to a frontier model.

Nemotron 3.5 Lightning runs on RTX PCs, DGX Spark and Station systems and Jetson devices, and is available on Hugging Face, ModelScope, OpenRouter and NVIDIA’s build platform; the model card lists an NVFP4 build. NeMo Switchyard is on GitHub with integration partners including LangChain, Kong and Boomi.

The benchmark and cost figures are NVIDIA’s own and not independently verified. Routing most calls to a cheaper model trades some quality for cost on the requests it keeps off the frontier model — a tuning choice each deployment has to make for itself.