Nvidia’s Nemotron 3 Ultra includes roughly 550 billion total parameters while operating with about 55 billion active parameters at any given moment. The model uses a mixture-of-experts design and incorporates Mamba-2 layers with standard Transformer attention to enable a one-million-token context window. Nvidia claims the Nemotron 3 Ultra delivers roughly 5x faster inference and about 30% lower costs compared with comparable open-weight alternatives.
The Nemotron 3 Ultra adopts a sophisticated mixture-of-experts approach that allows for efficient computational management. This design enables the model to utilize only a subset of its total 550 billion parameters, engaging 55 billion active parameters at any time to optimize processing efficiency.
Key to its architecture are the Mamba-2 layers, which play a critical role in supporting its advanced memory capabilities by facilitating a one-million-token context window. This configuration ensures that vast amounts of contextual information can be retained and processed simultaneously, enhancing the model’s ability to manage complex tasks.
Additionally, standard Transformer attention is integrated into the design, facilitating effective data management and improving processing speeds. The mixture-of-experts routing further optimizes the allocation of computational resources to diverse segments of the network, ensuring that the most relevant parts of the model are activated for any given task. By integrating these advanced features, Nvidia’s Nemotron 3 Ultra stands out as a technically sophisticated model designed to handle demanding processing requirements efficiently.
Artificial Analysis scored Nemotron 3 Ultra at 48 on its Intelligence Index. That score ranks Nemotron 3 Ultra above other American open-weight models, with Gemma 4 31B scoring 39, Nemotron 3 Super scoring 36, and OpenAI’s gpt-oss-120b scoring 33. Nemotron 3 Ultra leads the U.S. open-weight field by a comfortable margin compared with these alternatives. The listed index values show a clear numerical gap between Nemotron 3 Ultra and the next closest American options.
Nvidia claims Nemotron 3 Ultra is the top U.S. open-weight model by a comfortable margin. The available coverage states that Nemotron 3 Ultra tops every American open-weight AI system by a wide margin but still trails the Chinese-led frontier. The reporting therefore places Nemotron 3 Ultra as the highest-ranked American open-weight model while noting a relative gap with Chinese models. These rankings provide point-by-point comparisons among U.S. open-weight systems.
Nemotron 3 Ultra is the largest Nemotron 3 model to date. The Nemotron family is offered in Nano, Super, and Ultra sizes, with the first Nemotron-branded model released in November 2023 and the third generation announced in December 2025, making Ultra the largest member of that third generation. Nemotron 3 Ultra holds the leading position among American open-weight models by a comfortable margin while remaining behind Chinese-led competitors.
