Nvidia says Nemotron 3.5 Lightning can speed up enterprise AI agents

Nvidia has introduced Nemotron 3.5 Lightning, an open artificial intelligence model designed for high volume agent tasks, alongside NeMo Switchyard, a tool that routes tasks between different AI models.

Kyt Dotson reports for SiliconANGLE that the releases target companies that use AI agents for a mix of routine requests and more demanding reasoning work.

Nemotron 3.5 Lightning is a 30 billion parameter mixture of experts model. This approach activates only parts of a model for a given task, which can reduce computing demands. Nvidia says Lightning can generate output up to four times faster than comparable models and complete agentic tasks 30% faster. Those figures are company claims.

The company positions the model as an open, customizable option for organizations that want to adapt an AI system with their own data, tools and workflows. Nvidia says customers can post train the model using its NeMo software and their own hardware.

Routing tasks to different models

NeMo Switchyard addresses a separate challenge: selecting the appropriate model at each stage of an agent workflow. The open source library is intended to assess available models and send each prompt to one that best meets a chosen priority.

Developers can configure the routing strategy around quality, response time or cost. In practice, that could mean using a smaller, faster model for classification or summarisation, while reserving a more capable model for complex analysis, coding or specialist knowledge tasks.

Nvidia says it is also releasing post training datasets and training recipes for Lightning. The company is working with partners including Boomi, Cadence, Cognition, Kong, LangChain, Nous Research and Siemens on intelligent model routing.