Nvidia is preparing Nemotron 4, its next open-weight model family, built around a flagship version sized at one trillion parameters or more. The Decoder reported the plan, citing The Information, and pegged that figure at double the parameter count of Nemotron 3 Ultra, the version Nvidia shipped in June. Reaching that scale would put Nvidia’s flagship open model where two Chinese labs sit. A bigger parameter count does not automatically mean a stronger model.
Moonshot AI’s Kimi K3 carries 2.8 trillion parameters, well above where Nemotron 4 is aiming, and DeepSeek V4 Pro runs at 1.6 trillion, also past that mark. Parameter count sets a ceiling on how much a model can store, but training data quality, architecture choices, and post-training tuning determine how much of that capacity gets used. Nemotron 3 Ultra’s June debut already illustrated this: it led open US models on Artificial Analysis’s Intelligence Index, yet still trailed Kimi K2.6. On the index’s current scoring, Nemotron 3 Ultra sits at 38 points against roughly 60 for Kimi K3, a gap a doubled parameter count is unlikely to close on its own.
Nvidia is backing that ambition with money: it now expects to spend about $28 billion by 2031 training models in-house, three times its earlier plan, and Nemotron 4 could ship as soon as this fall. That timeline signals a multi-year commitment rather than a one-off release, and it puts a dollar figure on how seriously Nvidia is treating open-weight competition with Chinese labs.
A model this size is also expensive to run. One trillion parameters demands far more inference hardware and memory bandwidth than the models most enterprises currently deploy, and Nvidia would be selling both the chips and the reference model into that same market. That dual role gives Nvidia an incentive to grow parameter counts even where score gains are marginal, since larger open models drive more compute purchases.
The strategy also sits inside a policy fight. Nvidia has signed a petition opposing regulation of open models, filed while the Trump administration weighs narrower, model-specific bans on Chinese releases. A genuinely competitive Nemotron 4 would also put Nvidia head-to-head with OpenAI and other large customers who buy its chips, a rivalry the company has not addressed publicly.
Teams evaluating open-weight models this fall should benchmark Nemotron 4 against Kimi K3 and DeepSeek V4 Pro on Artificial Analysis’s Intelligence Index, not on parameter count alone. That index is where Nvidia’s current flagship still trails by more than 20 points.
The Decoder, which first reported this story, published its account on 12 August 2026.
