0
Illustration depicting the Nvidia logo, semiconductors, and a 3D-printed robotic hand model. / Reuters-Yonhap

Nvidia, which has long championed open-source AI software, released its own open-source AI model on the 11th (local time).

According to CNBC, Nvidia unveiled the lightweight AI model "Nemotron 3.5 Lightning" that day, introducing it as capable of running on a single GPU on a personal computer.

The model has 30 billion total parameters, but only 3 billion are activated during token processing. Other companies can download, use, and modify it for free, without needing Nvidia's permission or paying any fees.

Tech outlet Interesting Engineering explained that the model offers up to four times faster output speeds than similarly sized models, making it well-suited for the repetitive, iterative tasks handled by AI agents.

For example, it's designed to quickly handle processes like retrieving information, running a program, checking the results, and then directing the next task.

Nemotron 3.5 Lightning can be assigned simple, repetitive tasks while more complex judgment calls are handed off to larger AI models, saving on cost and processing time, the company said.

An Nvidia official said on the 11th that during development, the company used distillation — a technique that trains a model using data generated by a larger model — to transfer the performance of Nvidia's larger Nemotron models to this lightweight version.

Nvidia CEO Jensen Huang, together with other major tech industry figures, urged the U.S. administration in a social media post late last month to support open-source models, saying regulators should avoid hasty regulation that could push innovation overseas.

In an Axios interview published on the 22nd of last month, Huang also said free AI will greatly benefit the hardware industry, adding that free AI will also be good news for the semiconductor industry.

Tech outlet The Information reported on the 11th that Nvidia is developing an ultra-large AI model called "Nemotron 4," with 1 trillion parameters.

                                                                                                          Kim Hyun-min