Nvidia unveiled Nemotron 3 Ultra at Computex on Sunday, a 550B-parameter open-weight model (running 55B active parameters via a mixture-of-experts design) that CEO Jensen Huang positioned as the smartest open model built in America. Independent evaluator Artificial Analysis scored it 48 on its Intelligence Index, well ahead of the next U.S. open models, Google's Gemma 4 31B at 39 and OpenAI's gpt-oss-120b at 33, but behind China's Kimi K2.6 from Moonshot AI at 54. The model serves over 300 tokens per second, three to six times faster than Chinese rivals, and supports a 1-million-token context window. The launch is part of Nvidia's five-year, $26B plan to push open-weight AI as Chinese labs dominate the open ecosystem, rising from 1.2% of global open-model usage in late 2024 to roughly 30% by the end of 2025. Nemotron 3 Ultra ships June 4.
