Nvidia's Nemotron 4 to Be Twice as Large as Its Predecessor

Nvidia is working on Nemotron 4, a new open-weight model family intended to compete with the best openly available models worldwide. The Information reports that the largest model in the family will have at least one trillion parameters, twice as many as Nemotron 3 Ultra. Nvidia has tripled its cloud spending for own model training to $28 billion by 2031. A release could happen as early as autumn. This would bring Nvidia to a scale that Chinese vendors have long occupied: Moonshot AI's Kimi K3 has 2.8 trillion parameters, and Deepseek V4 Pro has 1.6 trillion. Nemotron 3 Ultra scored 48 points in the Artificial Analysis Intelligence Index, making it the strongest open US model, but behind Kimi K2.6 with 54; K3 now stands at 57. Nvidia is also a signatory to a petition against regulating open models, while the Trump administration considers targeted bans of certain Chinese models. The more companies host open models themselves, the more GPUs Nvidia sells. With Nemotron 4, however, Nvidia would also compete with its own major customers like OpenAI.