Nvidia released Nemotron 3.5 Lightning, a free open-source AI model built for enterprise use that runs on a single GPU. The company also launched NeMo Switchyard, a new model routing tool aimed at lowering costs on AI workloads.

This marks Nvidia’s first open-source model since CEO Jensen Huang posted an open letter on X in late July. In that letter he defended open AI models and urged the U.S. government to avoid restrictions that could push innovation overseas.

Nemotron 3.5 Lightning for Multi-Agent Systems

Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model. It targets high-volume, specialized tasks inside larger multi-agent systems. Nvidia reports the model delivers up to four times faster output speed and 30% faster agentic task completion than other models in its class.

The model is fully customizable. Organizations can post-train it on their own data. Kari Briski, Nvidia’s vice president of generative AI software, said the model is “remarkably easy to customize.” CodeRabbit used Nvidia’s standard training recipe to produce a router agent for $85 in about two hours. CrowdStrike and Harvey have also tested and customized the model.

For Nvidia, open-source AI supports a clear business goal. Huang told Axios last month that free AI should be great for hardware and great for chips.

NeMo Switchyard Routes Requests by Quality, Latency, and Cost

Alongside Lightning, Nvidia released NeMo Switchyard. This open-source model routing library automatically sends each AI agent request to the most suitable model based on quality, latency, and cost.

Internal benchmarks showed Switchyard kept frontier-level accuracy while cutting task completion cost to nearly one-third of the cost of using a single frontier model alone. Early partner results include:

  • Ramp matched frontier model performance while cutting costs by 58% and runtime by 33%.
  • LangChain achieved 74% lower cost by routing only 7% of calls to a frontier model.

Availability and Broader Context

Nemotron 3.5 Lightning is available on Hugging Face, ModelScope, OpenRouter, and Nvidia’s own platforms. Nvidia is already developing a next-generation Nemotron 4 model.

The release came one day after Meta Platforms CEO Mark Zuckerberg published a manifesto arguing for open-source AI, as Meta released a new coding model. It also follows Huang’s entry into a policy debate sparked by China’s Kimi K3 model, which raised concerns in Washington about intellectual property theft through distillation techniques.