Introduction
In a strategic announcement marking a prominent and bold shift in the company’s trajectory, Jensen Huang, CEO of Nvidia, unveiled the open-weight Nemotron 3 Ultra artificial intelligence model, featuring 550 billion parameters. This major revelation came during his keynote address at Computex 2026 in Taiwan’s capital, Taipei, on Monday. This advanced technical step represents the company’s boldest attempt to date to expand its influence beyond hardware manufacturing, breaking into the enterprise and major corporate AI software market, which places it in direct and intense competition with the world’s leading language model developers.
- Launch of the Nemotron model and its innovative architecture
- Superior performance and reduction of operational costs
- Building an integrated ecosystem for autonomous agent software
- Strengthening Nvidia’s competitive position in the software market
- FAQs
Launch of the Nemotron model and its innovative architecture
The new language model relies on an innovative software architecture known as mixture of experts, where only about 55 billion parameters are active per token processed, with a sparsity ratio of 90 percent. This precise architectural design makes it far more efficient and effective than its massive total parameter count suggests. According to evaluations by the independent agency Artificial Analysis, the model scored 48 on the agency’s intelligence index, placing it ahead and superior to all other available American open-weight models, including Google’s Gemma 4 which scored 39 points, although it still trails slightly behind China’s Kimi model which scored 54 points on the same index.
Superior performance and reduction of operational costs
Nvidia emphasized in its presentation that the Nemotron 3 Ultra model is reliably capable of generating more than 300 tokens per second, making it three to six times faster than competing models of similar intelligence levels such as DeepSeek and Moonshot, which typically process only between 50 to 100 tokens per second. The developing company claims this superior performance and exceptional speed reduces costs by about 30 percent for complex agent-based tasks compared to leading alternatives available in the market. This model has been meticulously fine-tuned to handle multi-step thinking and reasoning, autonomous planning, and self-correction over extended timeframes, which are essential capabilities for empowering AI agents.
Building an integrated ecosystem for autonomous agent software
Nemotron 3 Ultra proudly sits at the pinnacle of a new family of models featuring a mid-range version named Super, and the lightweight Nano Omni version that integrates vision, audio, and language for edge-embedded AI agents. Alongside the software models, Huang introduced the NeMo Clue framework specialized in planning and task delegation for agents, in addition to the Open Shell runtime layer focused on security, enterprise governance, and compliance. These advanced announcements came during the keynote address, which also included the unveiling of the Vera central processing unit, specifically designed for AI workloads, which Nvidia asserts offers twice the efficiency of traditional server chips, alongside the advanced RTX Spark chip.
Strengthening Nvidia’s competitive position in the software market
This wide and integrated set of announcements positions Nvidia not just as a market-dominant chipmaker, but as an integrated AI platform company competing fiercely with the likes of OpenAI, Google, and Meta in developing advanced language models, while maintaining its absolute dominance in hardware and infrastructure. The Vera platform has already begun attracting major cloud service providers including AWS, Google Cloud, and Microsoft. Computex 2026 provides an ideal global platform to reaffirm the company’s vision for future technological expansion and building a world entirely reliant on intelligent machines.
FAQs
Question: What is the total parameter count in the open Nemotron 3 Ultra model?
Answer: The model features 550 billion parameters and relies on the advanced mixture of experts architecture to activate a portion of them and improve efficiency.
Question: How does the speed of the newly developed model surpass market competitors?
Answer: It can process more than 300 tokens per second, making it three to six times faster and significantly reducing operational costs.
Question: What is Nvidia’s most prominent strategic goal in launching these models?
Answer: The company aims to establish its position as an integrated AI platform provider and expand its operations to include enterprise-focused software.
