Nvidia's Nemotron 4 aims for 1 trillion params, already trails rivals

Nvidia's next open-weight model will pack at least one trillion parameters, doubling the size of its predecessor. Yet by the time Nemotron 4 ships, that scale may look modest next to what Chinese labs are already shipping today.

According to reporting on Nvidia's roadmap, Nemotron 4's flagship version will be roughly twice as large as Nemotron 3 Ultra, which launched in June 2026. At the time of its release, Nemotron 3 Ultra held the title of strongest open US model on the Artificial Analysis Intelligence Index, a widely watched capability benchmark.

That leadership looks fragile against the current competitive field. DeepSeek's V4 Pro already runs at 1.6 trillion parameters, and Moonshot AI's Kimi K3 reaches 2.8 trillion, nearly three times what Nemotron 4 is targeting.

The scoreboard gap is just as stark as the parameter gap. Nemotron 3 Ultra sits at 38 points on the Artificial Analysis Intelligence Index, while Kimi K3 scores around 60.

Nemotron 4 won't arrive soon enough to close that gap quickly, either. The earliest plausible launch window is fall 2026, leaving Nvidia's open-weight flagship trailing rivals in both size and measured intelligence for well over a year after its predecessor debuted.

Betting big despite the gap

The scale-up isn't cheap. Nvidia has tripled its planned cloud spending on in-house model training, committing $28 billion through 2031 to build and run models like Nemotron.

That's a notable pivot for a company whose core business is selling the chips that power everyone else's models. Training frontier-scale systems in-house puts Nvidia in more direct competition with the very customers buying its GPUs, even as it keeps selling them the hardware to compete with it.

Nvidia's position on model openness is also part of the story. The company is among the signatories of a petition opposing new regulation of open-weight models, a stance that aligns with its strategy of positioning Nemotron as a freely available alternative to closed frontier systems from OpenAI, Anthropic, and Google.

The bigger question the Nemotron 4 numbers raise is whether raw parameter count still buys competitive standing. Kimi K3 outperforms Nemotron 3 Ultra by a wide margin despite the two projects converging on similar training approaches, suggesting that architecture, data quality, and post-training techniques may matter more than headline size going forward. Nvidia's trillion-parameter target may prove to be a floor rather than a ceiling by the time the model actually ships.