NVIDIA AI Releases Nemotron 3.5 Lightning and NeMo Switchyard

Long-running agents spend the vast majority of their operational time on high-volume execution tasks. Activities such as tool calls, result validation, and subagent delegation consume a massive portion of the available token budget, meaning that routing every single step through a frontier reasoning model introduces unnecessary costs and latency.

To solve this execution layer bottleneck, NVIDIA has developed Nemotron 3.5 Lightning as a 30B mixture-of-experts model featuring 3B active parameters. Built upon a hybrid architecture combining Mamba-2, MoE, and Attention, it supports a context length of up to 1 million tokens and was pre-trained on more than 20 trillion tokens using an NVFP4 recipe.

Stock photo for illustration only, not from the actual event

As the smallest member of the Nemotron 3 family, the model complements frontier counterparts like Nemotron 3 Ultra, which handle orchestration and planning while Lightning manages routine execution calls underneath. NVIDIA reports that the model delivers up to 4x output speed compared to similarly sized alternatives, achieving 86% accuracy on PinchBench while completing 10,000 tasks 30% faster than Qwen3.6 35B at comparable accuracy levels.

Complementing the model release, NeMo Switchyard serves as an open-source library designed to direct each step of an agent workflow to the most capable and efficient available model. It features tuning-free routers including an LLM classifier with session affinity, a stage router tracking recent tool activity, and an escalation router that begins with economical models and promotes tasks when sustained difficulty arises.

Source: MarkTechPost