NVIDIA has introduced Nemotron 3.5 Lightning, an open AI model built for fast, repeated tasks in long-running AI agents. The model aims to reduce the time and computing cost needed for tasks such as tool calls, result checks, and routine actions.
NVIDIA created Nemotron 3.5 Lightning as a smaller model for the execution side of AI agents. It has 30 billion parameters, with about 3 billion active parameters used for each token.
The model was developed by NVIDIA and was announced by Chris Alexiuk and Chintan Patel on the NVIDIA Developer Blog. NVIDIA says the model is part of its Nemotron family of open AI models.
NVIDIA’s Nemotron 3.5 Lightning uses a mixture-of-experts
Long-running AI agents can make many small calls while working on a task. Using a large reasoning model for every step can increase cost and slow down the system.
Nemotron 3.5 Lightning uses a mixture-of-experts, or MoE, design. A router sends each task to only a small part of the model, which helps reduce the computing work needed for each response.
The model also uses techniques such as speculative decoding to improve response speed. NVIDIA says it was trained with popular agent tools in mind, helping it handle repeated actions with lower delay.
Nemotron 3.5 Lightning can handle routine work such as checking tool results, running commands, formatting responses, and other high-volume tasks. NVIDIA says it can run on systems ranging from the DGX Spark to larger data centers, as well as local platforms such as Jetson and GeForce RTX systems.
NVIDIA also provides tools for adapting the model to specific jobs. Developers can fine-tune it and use routing tools such as NeMo Switchyard to send complex tasks to larger models while giving routine work to Lightning.
NVIDIA’s Nemotron 3.5 Lightning is designed mainly for execution tasks
Nemotron 3.5 Lightning is designed mainly for execution tasks, not every kind of AI problem. More complex planning and deep reasoning may still require larger frontier models, so the best setup can involve several models working together.
The main breakthrough is using a smaller open model for the high-volume work that AI agents perform repeatedly. This could help developers build AI agents that are faster and cheaper to run, while still using larger models when difficult reasoning is needed.
NVIDIA’s approach also points to a wider shift toward systems that route different tasks to different AI models. The NVIDIA Developer Blog says Nemotron 3.5 Lightning is available with its weights, data, and training recipes, giving developers more freedom to customize and deploy it.
