Nvidia is developing Nemotron 4, a new open-weight model intended to compete with top freely available models globally. The largest version of the model will have at least one trillion parameters, double the size of the previous Nemotron 3 Ultra. Nvidia has increased its cloud spending on in-house model training to $28 billion by 2031. The earliest possible release of Nemotron 4 is expected this fall. Even with one trillion parameters, Nvidia's model would still lag behind Chinese labs, which have already achieved higher parameter counts.
Nemotron 3 Ultra, the current strongest open US model, scored 38 points on the Artificial Analysis Intelligence Index at launch in June. However, it still trailed Kimi K2.6, which scored around 60 points. Kimi K3, released by Moonshot AI, has 2.8 trillion parameters, while DeepSeek V4 Pro has 1.6 trillion. Nvidia is also among the signatories of a petition against regulating open models, even as the Trump administration considers targeted bans on specific Chinese models.
Nvidia's strategy to self-host open models increases GPU sales, but Nemotron 4 could also put the company in direct competition with its major customers like OpenAI. The company's efforts reflect a broader industry trend of scaling up models to achieve better performance and market dominance. Source: thedecoder
