Nvidia CEO Jensen Huang unveiled the first open-source artificial intelligence (AI) model, “Nemotron 3.5 Lightning,” on Tuesday, August 11, sparking attention in the tech world. Less than 3 weeks ago, he publicly expressed support for open-source models. There is also news indicating that Nvidia is developing the “Nemotron 4” model, with parameter scale expected to reach 100 trillion.
Jensen Huang posted on the social media platform X on Tuesday, stating that “Nemotron 3.5 Lightning” possesses characteristics of intelligence, speed, efficiency, and openness.
According to Nvidia’s introduction, this is an open 300 billion parameter hybrid expert model designed for continuous online AI operation, enabling faster completion of large-scale, highly specialized tasks with output speeds up to 4 times that of models of similar scale.
This new Nemotron product is available for enterprises to download, use, and modify for free.
Nvidia stated that the model features lightweight characteristics, capable of running on a single GPU in personal computers.
On Tuesday, the AI chip giant also released “NeMo Switchyard,” an open-source model routing library designed to automatically allocate AI tasks to the most suitable models.
According to The Information, Nvidia is developing the “Nemotron 4.” Sources involved in the project revealed that developing “Nemotron 4” aims to compete with the world’s top open-source models.
Nvidia is one of the few large American companies that release open-source models. With increasing AI-related expenditures and low-cost Chinese models gradually nearing the capabilities of leading AI systems from American labs such as Anthropic and OpenAI, open-source models have attracted more attention this year.
Multiple employees involved in the project disclosed to The Information that the upcoming “Nemotron 4” model is expected to have at least 100 trillion parameters.
Nvidia has not confirmed the release date for “Nemotron 4,” and the final training is yet to be completed. However, employees involved in the project suggest that the model could potentially be ready as early as the end of this fall.
On July 24, Jensen Huang published an open letter supporting open-weight models. A few days later, Nvidia, along with Microsoft and other companies, established an AI Security Alliance dedicated to enhancing network security using open models and open-source software.
Open-source artificial intelligence is one of the hottest topics in Silicon Valley. On Monday, Meta CEO Mark Zuckerberg expressed support for open-source AI. Simultaneously, the company released a coding model called “Muse Spark.”
“Our goal should be to make America’s open-source models the best in the world,” Zuckerberg wrote.
(This article referenced reports from Reuters and CNBC.)