Nvidia, which recently led an open letter opposing regulations on open-source artificial intelligence, has released a lightweight AI model developed in-house. The company is also expected to introduce an open 1-trillion-parameter AI model.
On Aug. 11 local time, Nvidia unveiled Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts, or MoE, model.
Companies can download and use the model without licensing fees or prior approval. Nvidia has also released the model weights, allowing users to modify and fine-tune it freely.
Nvidia told CNBC that Nemotron 3.5 Lightning was developed using a technique known as distillation, drawing on its larger Nemotron models to deliver comparable performance in a smaller package.
Distillation is a method of training a new model on the responses generated by a larger AI model. Major AI companies have long used it to build more compact models. More recently, however, the practice has drawn scrutiny amid allegations that Chinese companies used it to replicate leading U.S. AI systems.
Nvidia said its development approach makes Nemotron 3.5 Lightning up to four times faster at generating tokens than comparable open models, while reducing total agent-task completion time by roughly 30%.
The company also introduced Nemo Switchyard, an open tool that identifies the complexity and requirements of a user request, then routes it to the most suitable AI model.
According to Nvidia, the tool can cut task costs to about one-third of the cost of running only the most powerful model, while maintaining comparable accuracy.
Separately, Reuters reported on the same day, citing technology outlet The Information, that Nvidia is quietly developing Nemotron 4, a next-generation AI model with 1 trillion parameters.
The model remains in training and does not yet have a confirmed release date. Sources said it could be completed as early as late this fall.
Kari Briski, Nvidia's vice president of generative AI, indirectly confirmed the company's next-generation-model plans in an emailed statement.
“Nvidia is investing in Nemotron because we believe every company and nation needs access to cutting-edge open models to strengthen safety and security, drive innovation, and provide a foundation that people can trust and rely on for generations,” Briski said.
Nvidia's advocacy for open AI models—and its decision to develop them directly—appears closely tied to the benefits that a broader open-model ecosystem could bring to its core businesses, including graphics processing units, or GPUs.
Unlike closed AI models from companies such as OpenAI, Anthropic, and Google, which often charge businesses substantial usage fees, open models are often free or available at relatively low cost. Because their weights are publicly available, users can also customize the models for specific needs.
A growing base of open-model users could pressure the developers of closed AI systems. For Nvidia, however, it could expand demand for AI chips while diversifying its customer base.
Nvidia CEO Jensen Huang recently told Axios that “free AI has to be good for hardware, semiconductors, and data centers.”
