NVIDIA has released two open-source artifacts designed for building always-on AI agents: the Nemotron 3.5 Lightning model and the NeMo Switchyard routing library. These tools address the high cost and latency of sending every step of long-running agent workflows to frontier reasoning models by providing specialized execution and routing capabilities.
Nemotron 3.5 Lightning is a 30B mixture-of-experts model with 3B active parameters, featuring a hybrid Mamba-2 + MoE + Attention architecture and a 1M-token context window. It achieves up to 4x faster output speed than similar-sized models and completes PinchBench tasks 30% faster than Qwen3.6 35B at comparable accuracy.
NeMo Switchyard is an open-source library that directs agent workflow steps to the most capable and efficient model available, offering tuning-free routers like escalation and stage routing. In benchmarks, it cut costs by 74% in a LangChain test by sending only 7% of calls to frontier models, with Cognition reporting lower mean costs for Devin Desktop.
Both artifacts are generally available under permissive licenses, with Lightning weights accessible on Hugging Face and ModelScope, enabling deployment on single GPUs like the DGX Spark or H100.