According to the latest data from the benchmarking platform Artificial Analysis, NVIDIA's newly launched open-source large model, Nemotron 3 Ultra, has been recognized as the highest-performing open-source AI model in the United States. However, it still falls short when compared to China's leading open-source models on a global scale. This model employs a Mixture of Experts (MoE) architecture, boasting a total parameter scale of around 550 billion, although only approximately 55 billion parameters are activated during actual operation. In the Artificial Analysis Intelligence Index evaluation, Nemotron 3 Ultra scored 48 points, outperforming Google's Gemma 4 31B, which scored 39 points, but it was surpassed by China's Kimi K2.6, which achieved a score of 54 points. Nevertheless, Nemotron 3 Ultra excels in terms of inference speed, significantly outpacing its peers in terms of tokens output per second, while also offering superior cost efficiency. NVIDIA highlights that this model is specifically tailored for code development, scientific research, and enterprise process agents. It delivers inference speeds up to five times faster than comparable open-source models, while also reducing usage costs by up to 30%.