Nvidia Introduces Cutting-Edge AI Model
Nvidia has launched its latest AI model, Nemotron 3 Ultra, during a keynote address at Computex on June 1. This powerful model features a staggering 550 billion parameters, marking it as the largest in the company’s Nemotron series and establishing it as a leading open-weight model in the U.S.
Performance and Ranking
Achieving an impressive output of over 300 tokens per second on a pre-release DeepInfra endpoint, Nemotron 3 Ultra surpasses its Chinese counterparts, which typically generate between 50 and 100 tokens per second. Despite its speed, it still trails Moonshot AI’s Kimi K2.6 in the intelligence rankings, receiving an Intelligence Index score of 48 compared to Kimi K2.6’s 54. This reflects a broader trend of Chinese models gaining ground in open-source AI.
Technical Architecture
The model utilizes a mixture-of-experts architecture, activating only 55 billion parameters at any given time to enhance efficiency. This approach allows Nvidia to achieve faster inference speeds and reduce costs by 30% compared to similar models. The Nemotron series is designed for various applications, with Ultra tailored for complex reasoning tasks, while the previous Super and Nano models cater to mid-range and simpler tasks, respectively.
