Nvidia unveils Nemotron 3 Ultra, its best open AI model

Nvidia takes a new step in open artificial intelligence with Nemotron 3 Ultra, its most powerful model to date. Presented at Computex 2026 in Taipei, this system impresses with its speed and reasoning capabilities. Yet, despite this major advance, Chinese labs continue to dominate the global open source AI rankings. Can Nvidia really catch up?

Nvidia signs a major advance in open source AI

On June 1st, in Taipei, Nvidia CEO Jensen Huang presented Nemotron 3 Ultra at the Computex show. With its 550 billion parameters, it is the largest open-weight model ever developed by the American company. This announcement also marks an important milestone for the open AI ecosystem in the United States.

In reality, only 55 billion parameters are activated simultaneously thanks to an architecture called “mixture of experts” (Mixture of Experts). This approach allows the model to mobilize only the necessary resources for each task. Result: faster execution and significantly reduced operating costs.

According to Nvidia, Nemotron 3 Ultra offers inference speeds up to five times higher than some competing models while reducing costs by about 30%. For companies developing autonomous agents or complex applications, this advantage could translate into significant productivity gains.

Independent assessments conducted by Artificial Analysis reinforce this impression. The model scores 48 on its intelligence index, which notably measures reasoning abilities, programming, general knowledge, and AI agent performance. This result places Nemotron 3 Ultra well ahead of other American open models, including those developed by Google or OpenAI.

This progress is all the more remarkable because its predecessor, Nemotron 3 Super, launched a few months earlier, scored 12 points lower. In a sector where every point gained becomes increasingly difficult to obtain, this improvement represents a real technological leap.

China maintains its lead in the open source race

Despite this progress, Nvidia has not yet managed to dethrone the Chinese leaders. The main obstacle has a name: Kimi K2.6, developed by Moonshot AI.

With a score of 54 on the Artificial Analysis index, Kimi K2.6 maintains a significant lead over Nemotron 3 Ultra. Even more impressively, the Chinese model is among the most powerful AI systems in the world, across all categories, whether open or proprietary.

This domination is not by chance. For several years, Chinese labs have multiplied the publication of high-level open source models. Conversely, American giants such as OpenAI, Anthropic, or Google generally reserve their most advanced technologies for proprietary services accessible via API.

This strategy has profoundly changed the market balance. At the end of 2024, Chinese models represented a marginal share of global open source AI usage. A year later, their share was already approaching 30%, illustrating the speed of their rise.

Aware of this strategic challenge, Nvidia invests massively in open AI. The group has already announced a $26 billion development plan over five years and is now working on Nemotron 4. This new generation will be developed with several major partners, including Mistral AI and Perplexity, within the Nemotron Coalition.

Nemotron 3 Ultra confirms that Nvidia is now the most ambitious American player in open source AI. Its lead in speed and efficiency could attract many companies.

However, the global open AI battle is no longer only between American giants. Today, China sets the pace of innovation and forces Western players to accelerate their efforts. For Nvidia, Nemotron 3 Ultra represents a major advance, but probably only a step in a technological competition that is just beginning.