Nvidia Releases Its Strongest Open AI Model Yet-but Still Trails China
Jensen Huang stepped onto the Computex stage in Taipei on Sunday, trademark leather jacket on his shoulders, to present what Nvidia now calls its most capable open AI system: Nemotron 3 Ultra. It is not just the company’s largest open model to date, but arguably the most advanced open‑weight AI model produced in the United States so far. Yet even this technological showpiece is still a step behind the newest generation of leading Chinese open‑weight models.
Nemotron 3 Ultra is built on an enormous architecture that boasts around 550 billion total parameters. In the world of large language models, parameters are the numerical “knobs” adjusted during training that encode patterns and knowledge. As a rule of thumb, more parameters mean a model can potentially capture richer structure in language and problem‑solving-though raw size alone is no longer the only metric that matters.
Despite its massive scale, Nemotron 3 Ultra doesn’t actually use all 550 billion parameters at the same time. Instead, the model employs a mixture‑of‑experts (MoE) design and activates only about 55 billion parameters for any given input. This approach is similar to running a gigantic hospital staffed with hundreds of specialized doctors, but calling in only the few experts suited to the particular patient in front of them. Most of the “experts” remain idle on each request, which is exactly what makes the system both huge in capability and efficient in practice.
This MoE architecture is central to how Nvidia squeezes high performance out of a model that would otherwise be prohibitively expensive to run at full capacity. Rather than forcing every part of the network to process each query, a router component dynamically selects which subset of experts should handle a given prompt. The result is that the model can exhibit the breadth of a 550‑billion‑parameter system while incurring the computational cost closer to that of a far smaller dense model.
From a purely American perspective, Nemotron 3 Ultra represents a milestone. Benchmarks shared by Nvidia indicate that it outperforms other U.S.‑developed open‑weight models across a wide spectrum of tasks, including coding, reasoning, and general language understanding. For companies, researchers, and developers looking for a powerful model they can study, self‑host, and customize without relying solely on closed black‑box systems, this is a significant upgrade.
However, the global race in open‑weight AI is no longer just a story of U.S. tech champions. In China, several frontier models-both closed and open‑weight-have recently leapt ahead on standard evaluation suites. While exact scores vary by benchmark and task, Chinese‑led open‑weight systems increasingly demonstrate stronger performance in complex reasoning, multilingual capabilities, and math‑heavy workloads than their American counterparts, including Nemotron 3 Ultra.
The gap is not only about accuracy scores. Chinese research groups and companies have been extremely aggressive in scaling up training runs, experimenting with highly optimized MoE designs of their own, and releasing models that can be fine‑tuned cheaply on commodity hardware. Combined with heavy government support and a dense local ecosystem of AI startups, this has pushed China’s open‑weight frontier a notch ahead of what U.S. firms have so far placed in the public domain.
Nemotron 3 Ultra should therefore be seen as both an achievement and a warning sign for the U.S. AI ecosystem. On one hand, Nvidia has proven that it can translate its dominance in AI hardware-GPUs, networking, and systems-into a serious contender on the open‑model side. On the other hand, the fact that its best open‑weight model still trails Chinese leaders underscores how quickly the international balance of power in AI research and deployment is shifting.
For Nvidia’s customers, the model serves a strategic purpose. By offering an open‑weight system that is deeply optimized for its own chips and infrastructure, Nvidia encourages enterprises to build directly on its stack. Companies can host Nemotron 3 Ultra in their own data centers, tune it with proprietary data, and integrate it into existing workflows while retaining more control than with closed, API‑only systems. That combination of power, flexibility, and infrastructure lock‑in is precisely what Nvidia is betting on.
Technically, the choice of MoE also signals where the broader industry is heading. The early days of language models were about ever‑denser networks-piling more and more parameters into a single monolithic structure. Today, as training costs and energy consumption soar, architects are converging on sparse designs like Nemotron 3 Ultra’s. These systems can scale parameter counts into the hundreds of billions or even trillions while keeping inference latency and cost within commercial bounds.
At the same time, performance leadership is only one dimension. Open‑weight models come with trade‑offs in safety, controllability, and misuse. More powerful open systems mean that not only large corporations but also smaller groups gain access to near‑frontier capabilities. Countries, including the U.S. and China, are beginning to grapple with how to regulate such models without choking off innovation. Nvidia’s move adds another heavyweight system to the open landscape, making that policy conversation even more pressing.
For developers and researchers, Nemotron 3 Ultra is likely to become a new default reference point. Many will compare it against both American open models and top Chinese releases to evaluate where each ecosystem stands in coding, scientific reasoning, creative writing, and multilingual use cases. As more independent benchmarks and real‑world applications emerge, the exact contours of Nvidia’s performance gap with Chinese leaders will become clearer.
Looking forward, Nemotron 3 Ultra is almost certainly not the endpoint for Nvidia. The company has every incentive to iterate rapidly: extending the Nemotron family, training even larger or more specialized MoE systems, and tightening the integration between its models and GPU platforms. Whether that will be enough to close the distance with the fastest‑moving labs in China remains an open question-but the direction of travel is obvious. The AI race is no longer just about who can ship the biggest chip; it is increasingly about who can translate that hardware edge into the most capable, widely usable, and openly accessible models.