NVIDIA has released two open pieces of the agent stack: a small fast model for the busy work inside agent runs, and a router that picks the right model for each step.
Nemotron 3.5 Lightning packs 30B parameters but activates only 3B per token, using a Mamba-2, MoE, and attention hybrid with a one-million-token context window. On output speed the vendor claims a 4x edge over similarly sized models, and a 30 percent gain on 10,000 PinchBench tasks relative to Qwen3.6 35B at similar accuracy. A single GPU, from a DGX Spark to an H100, can serve it.
Lightning is aimed at the execution layer of long-running agents: tool calls, result validation, subagent delegation, and review routing, where routing every step to a frontier reasoning model wastes time and money. Firms across cybersecurity, legal, coding, finance, and healthcare are adapting it, including CrowdStrike, Harvey, CodeRabbit, Fastino Labs, and Lila Sciences. The model is out now under the permissive OpenMDW-1.1 license, with open weights, training data, and recipes shipped alongside.
NeMo Switchyard, its companion, routes each workflow step to the most capable and efficient model using tuning-free routers, including one built on an LLM classifier. NVIDIA says internal tests cut task costs to a third. Both tools anchor an August campaign pushing local AI, joined by the open-weights LTX-2.5 world model.
