Nvidia Nemotron 3 Ultra: A New Era for Open AI
- Nvidia unveiled Nemotron 3 Ultra at Computex.
- The model features 550 billion total parameters (55 billion active).
- Ranked as the smartest US open-weight model with an Intelligence Index of 48.
- Boasts 5x faster inference and 30% lower costs.
- Aims to compete with Chinese leaders like Kimi K2.6.
Nvidia Unveils Nemotron 3 Ultra: A Bold Leap in Open AI
At the Computex stage in Taipei, Jensen Huang, NVIDIA‘s CEO, introduced Nemotron 3 Ultra – the company’s most expansive and, for now, the most intelligent open-weight artificial intelligence model developed in America. This model, packing roughly 550 billion total parameters, represents a significant stride in Nvidia‘s efforts to bolster the US position in the global AI race.
“Nemotron 3 Ultra showcases our unwavering commitment to open AI innovation. We’re not just building powerful tools; we’re shaping a future where advanced technology is accessible to all developers,” stated an Nvidia spokesperson, emphasizing the strategic importance of the new model.
Despite its impressive capabilities, the model faces stiff competition from Chinese advancements, which continue to lead in certain key performance metrics.
The Architecture Behind the Power
At the core of Nemotron 3 Ultra is an innovative “mixture-of-experts” design, allowing the model to operate on only 55 billion active parameters at any given moment. This approach drastically reduces operational costs while maintaining high performance. Nvidia claims 5x faster inference and 30% lower costs compared to comparable open-weight alternatives.
The model also incorporates a hybrid architecture combining Mamba-2 layers, standard Transformer attention, and mixture-of-experts routing. Mamba-2 optimizes processing for long sequences, enabling Nemotron 3 Ultra to support a 1-million-token context window. This means an AI agent can, in theory, process entire large codebases or hundreds of research documents simultaneously.
Key Technologies Powering Nemotron 3 Ultra
- Mixture-of-Experts: Efficient parameter utilization.
- Mamba-2: Optimized long-sequence processing.
- Multi-Token Prediction (MTP): Accelerated text generation.
- Reinforcement Learning: Ability to plan and execute multi-step tasks.
The Intelligence Race: US vs. China
Independent evaluator Artificial Analysis, which partnered with Nvidia, scored 48 on its Intelligence Index. This positions it as the top US open-weight model, comfortably surpassing Google‘s Gemma 4 31B at 39 and OpenAI‘s gpt-oss-120b at 33.
“Nemotron 3 Ultra sets a new benchmark for open models in the US, showcasing significant advancements in reasoning and coding capabilities,” noted analysts at Artificial Analysis.
However, on the global stage, Nemotron 3 Ultra still trails some Chinese competitors. Moonshot AI‘s Kimi K2.6, released in April 2026, scored 54, ranking fourth among all AI models globally, both closed and open. This 6-point gap highlights a meaningful difference in intelligence capabilities from Chinese developments.
Speed and Accessibility
Where Nemotron 3 Ultra truly shines is its inference speed. On a pre-release DeepInfra endpoint, the model served over 300 output tokens per second. In contrast, Chinese models in its intelligence class, such as DeepSeek V4 Pro and Kimi K2.6, typically serve at 50–100 tokens per second through their commercial APIs. This speed advantage is critical for real-world deployments, especially for autonomous agents executing long, multi-step tasks where waiting for each step compounds quickly.
While running a 550-billion-parameter model typically requires datacenter-level hardware, Nvidia makes it accessible through its API or cloud providers, allowing developers to leverage its power without owning the underlying supercomputer hardware.
Nvidia’s Strategic Investment in Open AI
Nvidia is actively working to reshape the open AI landscape. The company has publicly disclosed a five-year plan to invest 26 billion dollars in open-weight AI development. Nemotron 3 Ultra is the most visible outcome of this significant bet so far.
Nvidia is already developing Nemotron 4 through the Nemotron Coalition, a group of eight AI labs including Mistral AI and Perplexity, assembled to co-develop open frontier models on DGX Cloud infrastructure.
“We see immense potential in open AI and believe that collaborating with leading labs will accelerate innovation. The Nemotron Coalition is our answer to the need for a more open and accessible AI ecosystem,” commented an Nvidia representative on the company’s strategy.
The release of Nemotron 3 Ultra on June 4, 2026, marks a pivotal moment in this strategy, demonstrating Nvidia‘s commitment to advancing open AI technologies and competing on a global scale.
Frequently Asked Questions (FAQ)
-
What is Nemotron 3 Ultra?
Nemotron 3 Ultra is Nvidia‘s largest and most intelligent open-weight artificial intelligence model developed in the US. It utilizes a “mixture-of-experts” architecture and features 550 billion parameters.
-
How does Nemotron 3 Ultra compare to Chinese models?
In terms of inference speed, Nemotron 3 Ultra significantly outperforms Chinese models, delivering over 300 tokens per second. However, in overall intelligence index, it currently trails leaders like Moonshot AI‘s Kimi K2.6.
-
What are the key features of Nemotron 3 Ultra?
Key features include its mixture-of-experts architecture, a 1-million-token context window, Mamba-2 layers for efficient long-sequence processing, and multi-token prediction for accelerated generation.
-
How can I access Nemotron 3 Ultra?
While running the model requires significant computational power, Nemotron 3 Ultra can be accessed via Nvidia‘s API or through cloud providers.
