Sources: Nvidia is developing a Nemotron 4 model with 1T+ parameters, up from Nemotron 3 Ultra's 550B parameters but smaller than leading Chinese open models
The Information:
Context & Ripple Effects
Nvidia has moved its open-weight Nemotron line from the hybrid-MoE 120B Nemotron 3 Super to 550B Nemotron 3 Ultra, which related coverage described as the strongest open U.S. model while still behind Kimi K2.6. The reported Nemotron 4 effort raises the scale target again rather than marking a one-off model experiment.
Nvidia also released a smaller 30B Nemotron model and an open agent-routing library on the same day, indicating that the company is maintaining both large frontier candidates and lighter models for deployment.
First-order effects
- Nvidia’s next Nemotron flagship is being positioned above Nemotron 3 Ultra’s 550B-parameter scale, intensifying its effort to narrow the gap with leading Chinese open models.
- Developers tracking Nvidia’s open-model roadmap now face a broader prospective Nemotron range: a lightweight 30B option, Ultra at 550B, and a reported trillion-parameter successor.
Second-order effects
- Kimi K2.6 and other leading Chinese open-model developers face a more direct U.S. open-weight challenger at the top end, rather than competing only with Nvidia’s 550B Ultra.
- Nvidia’s model-routing software becomes more relevant if customers must choose among markedly different Nemotron sizes, shifting part of model selection toward routing and inference operations.
Third-order effects
- If Nvidia continues pairing open models at multiple scales with deployment tooling, competition in open AI will increasingly center on complete model portfolios rather than a single benchmark-leading release.
- The reported scale-up reinforces compute-capacity economics as a dividing line in open models: only providers able to support both frontier-scale training and deployable smaller models can cover the full stack.
The trend: Open-model competition is becoming a portfolio-and-infrastructure contest, with Nvidia expanding both the upper scale of its flagship models and the tooling around their deployment.