NVIDIA releases Nemotron 3.5 Lightning for agent execution

NVIDIA has released Nemotron 3.5 Lightning, an open model designed to handle routine execution in AI agent systems, including tool calls, output checks and result formatting. The company is pairing it with NeMo Switchyard, an open source routing library that can send complex planning or reasoning requests to larger models while directing repetitive work to Lightning.

Nemotron 3.5 Lightning is a 30B mixture-of-experts model with 3B active parameters at a time. NVIDIA says the design gives the model some of the capacity of a larger system while keeping computing demands closer to a smaller one. The model has been trained for agent harnesses including OpenClaw and Hermes Agent, and NVIDIA says it produces output up to four times faster than similarly sized models.

Company-published benchmark results place Lightning on the accuracy-speed frontier for small open models in the Artificial Analysis Intelligence Index. On PinchBench, NVIDIA says the model reached 86% accuracy and completed 10,000 tasks 30% faster than Qwen3.6 35B at similar accuracy, though the release does not cite extended production results from named users.

NVIDIA has released model weights, training data and post-training recipes under the OpenMDW-1.1 license. Developers can test the model through build.nvidia.com and OpenRouter, access weights via Hugging Face and ModelScope, and deploy it with guides for vLLM, SGLang and TensorRT-LLM.