NVIDIA Nemotron 3 Ultra: 550 billion parameters, but China is still ahead

NVIDIA Unveils Nemotron 3 Ultra: America's Largest and Smartest Open-Weight Model

On June 1, 2026, during the Computex conference in Taipei, NVIDIA unveiled Nemotron 3 Ultra, its largest and most powerful open-weight artificial intelligence model ever created. With a total of 550 billion parameters and 55 billion active parameters, this model represents a significant leap forward in the global competition for the development of artificial intelligence models.

Architecture and Performance

Nemotron 3 Ultra uses a "mixture-of-experts" architecture, a design that allows only the relevant parameters for a specific request to be activated, reducing operational costs and improving efficiency. This approach enables the model to perform inferences up to five times faster than Chinese rivals and with 30% lower operational costs than comparable open-weight alternatives.

The model was evaluated by Artificial Analysis, an independent entity that awarded Nemotron 3 Ultra a score of 48 on the Intelligence Index, a composite benchmark that aggregates 10 evaluations ranging from reasoning, programming, general knowledge, and agentic performance. This score positions Nemotron 3 Ultra as the smartest open-weight model produced in the United States, far surpassing national competitors like Google's Gemma 4 31B and OpenAI's gpt-oss-120b.

The Nemotron Family

NVIDIA announced the Nemotron family in November 2023, with the launch of the third generation in December 2025. The Nemotron family includes three models of different sizes: Nano for light tasks, Super for mid-level business applications, and Ultra for complex reasoning workloads. All three models share the same hybrid architecture that combines Mamba-2 layers, standard Transformer attention, and mixture-of-experts routing.

Mamba-2 is an alternative to standard attention that processes long sequences at a fraction of the cost, relevant for models capable of keeping a million tokens in memory simultaneously. Nemotron 3 Ultra supports a context window of 1 million tokens, theoretically allowing an agent to have an entire large codebase or hundreds of research documents at its disposal simultaneously.

The Ultra model also includes a technique called multi-token prediction (MTP), which allows the model to predict several future tokens simultaneously instead of one at a time, accelerating generation. All three Nemotron 3 models have been post-trained using reinforcement learning in multiple interactive environments, teaching them to plan and execute multi-step tasks rather than just answering questions.

Accessibility and Speed

Nemotron 3 Ultra is publicly available, and its training recipes will be made public. To run it, a supercomputer is essential, but it can be accessed via NVIDIA's API or cloud service providers without owning the necessary hardware, similar to using models like GPT or Claude through a browser.

The speed of Nemotron 3 Ultra is one of its strengths. On a pre-release DeepInfra endpoint, the model served over 300 output tokens per second. Chinese models in the same intelligence class, such as DeepSeek V4 Pro and Kimi K2.6, were served at 50–100 tokens per second through their current commercial APIs. This difference in speed is crucial for practical implementations, especially for autonomous agents performing complex, multi-step tasks where waiting for each step quickly adds up.

The Global Competition

Despite its impressive performance, Nemotron 3 Ultra fails to surpass Chinese models in terms of intelligence. For example, Moonshot AI's Kimi K2.6 has a score of 54 on the Intelligence Index, six points higher than Nemotron 3 Ultra. This gap represents a significant difference, positioning Kimi K2.6 in fourth place among all global artificial intelligence models, both open and closed, just three points behind the leading proprietary models from Anthropic, Google, and OpenAI.

The current situation reflects a trend where Chinese labs are flooding the open ecosystem with strong models, while American companies like OpenAI, Anthropic, and Google keep their best systems behind APIs. NVIDIA is the largest American name actively seeking to reverse this trend, with a five-year plan to spend $26 billion on the development of open-weight artificial intelligence models.

The Future of Nemotron

Nemotron 3 Ultra is the most visible result of this investment so far. NVIDIA has also announced that it is already working on Nemotron 4, the next generation, developed through the Nemotron Coalition, a group of eight artificial intelligence labs, including Mistral AI and Perplexity, assembled by NVIDIA in March 2026 to co-develop open frontier models on DGX Cloud infrastructure. Nemotron 3 Ultra will be available from June 4, 2026.