Jensen Huang, CEO of Nvidia, commanded the stage at Computex in Taipei on Sunday, dressed in his signature leather jacket, to introduce Nemotron 3 Ultra. This latest offering marks Nvidia’s largest open AI model to date, boasting approximately 550 billion total parameters and establishing itself, at least for the present, as the most intelligent open-weight model developed in the United States. While representing a significant leap forward in American AI capabilities, independent evaluations indicate that Nemotron 3 Ultra, despite its advancements, has not yet surpassed the leading models emanating from China in terms of raw intelligence.
This unveiling at Computex, a pivotal global event for the information and communication technology industry held annually in Taiwan, underscored Nvidia’s accelerating commitment to the open-source AI ecosystem. Huang’s keynote address, a highlight of the conference, typically sets the tone for future technological trends, and this year, the focus was firmly on democratizing advanced AI. Nemotron 3 Ultra leverages a sophisticated design known as Mixture-of-Experts (MoE), which allows the model to operate with only about 55 billion active parameters at any given moment. This innovative approach optimizes computational efficiency, allowing for a broader knowledge base (indicated by the higher total parameter count) while maintaining practical operational costs. Parameters are fundamental to an AI model’s capacity for knowledge and understanding, with a greater number generally correlating with enhanced performance and broader capabilities.
The Strategic Imperative: Nvidia’s Push in Open-Weight AI
Nvidia’s introduction of Nemotron 3 Ultra is more than just a product launch; it’s a strategic maneuver in the increasingly competitive global landscape of artificial intelligence. The company has explicitly committed to a five-year plan involving a substantial $26 billion investment in open-weight AI development, a clear signal of its intent to shape the future of accessible, high-performance AI. This commitment comes amidst a dynamic environment where the U.S. faces significant competition, particularly from Chinese AI labs that have been rapidly advancing their open-source offerings.
The distinction between "open-weight" and "closed" or "proprietary" models is crucial. Open-weight models have their underlying parameters and architecture publicly available, allowing developers to inspect, modify, and build upon them. This fosters innovation, transparency, and a wider developer community. In contrast, closed models, such as many of the flagship systems from OpenAI, Anthropic, and Google, keep their core architecture and weights proprietary, accessible only through APIs. While closed models often represent the cutting edge of AI intelligence, their proprietary nature limits external scrutiny and customization. Nvidia’s strategy, therefore, aims to bridge the gap by offering top-tier performance within an open framework, thereby empowering a broader base of developers and enterprises.
Understanding the Mixture-of-Experts Architecture
At the heart of Nemotron 3 Ultra’s efficiency lies its Mixture-of-Experts (MoE) architecture. To conceptualize this, imagine a vast hospital housing hundreds of highly specialized medical professionals. When a patient presents with a specific ailment, only the relevant specialists—say, a cardiologist for heart issues or a neurologist for brain conditions—are called upon, rather than the entire medical staff. This selective engagement is precisely how MoE models function. Instead of activating all 550 billion parameters for every single query, the model intelligently routes incoming requests to the most appropriate "expert" sub-networks, each comprising a portion of the total parameters.
This intelligent routing mechanism yields substantial benefits:
- Reduced Inference Costs: By only activating a fraction of the total parameters, the computational resources required for each inference (generating an output from an input) are significantly lowered. Nvidia reports that this design enables 5x faster inference speeds and reduces operational costs by 30% compared to other comparable open-weight alternatives.
- Scalability: MoE allows for the creation of extremely large models (in terms of total parameters) without the prohibitive runtime costs that would typically accompany such scale in traditional "dense" models, where all parameters are active during inference.
- Specialization: Different expert networks can specialize in distinct types of knowledge or tasks, potentially leading to more nuanced and accurate responses across a wider range of domains.
This architectural choice is particularly vital for applications requiring high throughput and cost-effectiveness, such as autonomous AI agents and large-scale enterprise deployments.
Performance Benchmarks and the Intelligence Index
Independent evaluation is paramount in assessing AI model capabilities. Artificial Analysis, a respected evaluator in the AI space, partnered with Nvidia for a pre-release assessment of Nemotron 3 Ultra. Their findings placed the model at an impressive 48 on their Intelligence Index. This composite benchmark rigorously aggregates scores from 10 distinct evaluations, spanning critical AI competencies such as reasoning, coding proficiency, general knowledge, and agentic performance—the ability of an AI to plan and execute multi-step tasks. A higher score on this numbered scale denotes greater intelligence.
This score of 48 firmly establishes Nemotron 3 Ultra as the leading U.S. open-weight model, surpassing its closest American counterparts by a comfortable margin. For context, Google’s Gemma 4 31B registered a score of 39, Nvidia’s own Nemotron 3 Super achieved 36, and OpenAI’s gpt-oss-120b scored 33. The advancement over its predecessor, Nemotron 3 Super (released in March 2026 with 120 billion parameters), is particularly striking, demonstrating a 12-point jump on the Intelligence Index. This leap is considered significant in the rapidly evolving benchmarking landscape, indicating substantial improvements in capabilities. Nemotron 3 Super was already regarded as a robust open model suitable for autonomous agents, making Ultra’s enhanced performance a testament to Nvidia’s accelerated development.
The Nemotron Family: A Unified Ecosystem
Nvidia’s foray into the model business predates common perception, with the first Nemotron-branded model launching in November 2023. The third generation of the family was subsequently announced in December 2025, laying the groundwork for the Ultra model’s release. The Nemotron family is designed with a tiered approach to cater to diverse computational needs and application complexities:
- Nemotron Nano: Tailored for lightweight tasks and edge deployments, where computational resources are limited but basic AI capabilities are required.
- Nemotron Super: Designed for mid-range enterprise applications, offering a balance of performance and efficiency for more substantial workloads.
- Nemotron Ultra: The flagship model, engineered for complex reasoning workloads and advanced AI agentic performance, demanding significant computational power.
All three models within the Nemotron 3 series share a sophisticated hybrid architecture that combines Mamba-2 layers, standard Transformer attention mechanisms, and the aforementioned Mixture-of-Experts routing. This blend of technologies is crucial for their performance and efficiency. Mamba-2, in particular, offers an alternative to traditional Transformer attention, excelling at processing extremely long sequences of data at a fraction of the computational cost. This is highly relevant for models designed to maintain vast amounts of information in memory simultaneously. Nemotron 3 Ultra, for instance, supports an impressive 1-million-token context window, theoretically allowing an AI agent to process an entire large codebase, hundreds of research documents, or extensive conversation histories concurrently.
Furthermore, the Ultra model incorporates a technique called multi-token prediction (MTP), which significantly accelerates text generation by enabling the model to predict several future tokens at once, rather than the conventional one-token-at-a-time approach. All Nemotron 3 models underwent post-training using reinforcement learning across multiple interactive environments. This advanced training methodology teaches the models to plan and execute multi-step tasks autonomously, moving beyond mere question-answering to genuinely agentic behavior.
While Nemotron 3 Ultra’s weights and training recipes are being made public, running a 550-billion-parameter model still necessitates datacenter-grade infrastructure. However, Nvidia mitigates this barrier to entry by offering access through its API and various cloud providers, mirroring the accessibility models like GPT or Claude provide via web browsers. This approach democratizes access to powerful AI, even for those without direct access to supercomputing hardware.
The Global Intelligence Contest: Speed vs. Raw Brainpower
Despite Nemotron 3 Ultra’s formidable capabilities, particularly its speed, the global AI intelligence contest reveals a more nuanced picture. On a pre-release DeepInfra endpoint, Nemotron 3 Ultra demonstrated exceptional speed, serving over 300 output tokens per second. This starkly contrasts with leading Chinese models in its intelligence class, such as DeepSeek V4 Pro and Kimi K2.6, which typically operate at 50-100 tokens per second through their commercial APIs. This speed differential is a critical advantage for real-world deployments, especially for autonomous agents engaged in lengthy, multi-step tasks where cumulative wait times can significantly impact efficiency and user experience.
However, raw processing speed is but one facet of AI performance. The Artificial Analysis chart also plainly illustrates the intelligence gap. While Nemotron 3 Ultra scores 48 on the Intelligence Index, China’s Kimi K2.6 from Moonshot AI achieves a higher score of 54. This six-point difference on the index represents a meaningful disparity in overall intelligence. Kimi K2.6, released in April 2026, currently ranks as the fourth most intelligent AI model globally, whether closed or open-source. It trails only three points behind the proprietary flagship models from Anthropic, Google, and OpenAI, which are currently tied at 57.
This situation reflects a broader trend in the U.S. open-weight AI landscape. While American tech giants like OpenAI, Anthropic, and Google primarily keep their most advanced systems proprietary and behind APIs, Chinese laboratories have been actively and successfully contributing robust models to the open ecosystem. This strategic divergence has allowed Chinese open-source models to dramatically increase their global usage, surging from approximately 1.2% in late 2024 to around 30% by the end of 2025, as reported by Decrypt in March. Nvidia, through initiatives like Nemotron 3 Ultra and its significant financial commitment, is arguably the most prominent American entity striving to reverse this trend and bolster the U.S. position in open-source AI.
Future Outlook: Nemotron 4 and the Coalition for Open AI
Nvidia’s ambition doesn’t stop with Nemotron 3 Ultra. The company has already announced that development is underway for Nemotron 4, the next generation of its open-weight models. This future iteration is being developed through the Nemotron Coalition, an alliance formed by Nvidia in March 2026. This coalition brings together eight prominent AI labs, including leading players like Mistral AI and Perplexity, to collaboratively develop frontier open models leveraging Nvidia’s DGX Cloud infrastructure. This collaborative approach underscores Nvidia’s strategy to pool expertise and resources to accelerate the pace of innovation in the open-source AI domain, directly challenging the dominance of closed, proprietary systems and aiming to push the boundaries of what open-weight models can achieve.
The release of Nemotron 3 Ultra on June 4, 2026, marks a pivotal moment for Nvidia and the broader open-source AI community. It demonstrates that cutting-edge performance, efficiency, and advanced capabilities can be delivered within an open framework, fostering greater collaboration and accessibility. While the intelligence gap with certain Chinese models remains a challenge, Nvidia’s continuous investment, technological innovations like MoE and Mamba-2, and strategic partnerships through the Nemotron Coalition position the U.S. to significantly strengthen its standing in the global AI race, particularly in the critical open-weight sector. The ongoing competition promises to drive further advancements, ultimately benefiting developers, enterprises, and the evolution of AI worldwide.
Concluding Analysis: A Balanced Perspective
Nemotron 3 Ultra represents a compelling advancement for American open-weight AI. Its blend of high parameter count, MoE efficiency, impressive inference speed, and substantial context window makes it highly attractive for enterprise adoption and the development of sophisticated AI agents. The 12-point jump in intelligence over its predecessor highlights Nvidia’s rapid development cycle and commitment to innovation.
However, the continued lead of certain Chinese open-weight models in raw intelligence, as evidenced by Kimi K2.6’s score, serves as a clear reminder of the intense global competition. This gap underscores the strategic importance of sustained investment and collaborative efforts, such as the Nemotron Coalition, to ensure the U.S. remains at the forefront of AI development across all dimensions—speed, cost-efficiency, and ultimate intelligence.
The democratizing effect of powerful, open-weight models cannot be overstated. By making these advanced tools more accessible and affordable to run, Nvidia is not only enhancing its own market position but also contributing to a more diverse and innovative AI ecosystem globally. The strategic implications extend beyond commercial competition, touching upon national security, economic competitiveness, and the future trajectory of technological leadership. As the AI landscape continues to evolve at an unprecedented pace, Nemotron 3 Ultra stands as a significant milestone, setting a new benchmark for what open-source AI can achieve from American shores.
