Nvidia releases next major AI models

Nvidia presents Nemotron 3 Ultra as its new frontier model, and Cosmos 3, a world model for robotics and autonomous systems.

listen Print view
Nvidia logo on building facade, treetops in the foreground

(Image: Michael Vi/Shutterstock.com)

3 min. read
By
  • Carolin Riethmüller

Nvidia has unveiled several new AI models that are causing a stir. Among them is Nemotron 3 Ultra, the most powerful model in the new Nemotron-3 family.

According to Nvidia, the model offers five times faster inference and up to 30 percent lower costs compared to previous versions. Support for a context window of one million tokens allows for processing large amounts of data. This makes it, like most currently published top models, particularly suitable for long and complex tasks.

With Cosmos 3, Nvidia is expanding its range of world models for physical AI applications with a “world's first fully open omnimodel”. Previously, developers had to work with separate models for different functions such as world generation (Cosmos Predict), scene understanding (Cosmos Reason), and rule set generation (Cosmos Policy).

According to the company, Cosmos 3 was trained on vast amounts of multimodal data and can, for example, simulate scenarios before machines like robots or autonomous driving systems actually act – taking into account physics, among other factors. Cosmos 3 includes Cosmos 3 Super, Nano, and soon Edge, each with different user focuses.

Many AI companies are currently investing in this area of “world models.” While language models like ChatGPT can communicate impressively, they lack an understanding of physical relationships. The fact that they write that a ball falls when released is more because they have read it frequently and reproduced it, not because they truly understand gravity. This is intended to be different with world models.

Industry observers therefore see world models as an important prerequisite for powerful robotics, autonomous systems, and, in the long term, even more general forms of artificial intelligence. Google researchers see world models similarly: with enormous potential for robotics and autonomous systems. However, compared to language models, they are still in the year 2021.

Videos by heise

Perhaps that's why Nvidia has now launched an Nvidia Cosmos Coalition. According to the company, participants in this cooperation include Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI.

However, there was also news from the American company in the hardware sector at Computex: Nvidia is working in parallel on its own hardware roadmap: The company plans two more processor generations for PCs by 2030.

(rie)

Don't miss any news – follow us on Facebook, LinkedIn or Mastodon.

This article was originally published in German. It was translated with technical assistance and editorially reviewed before publication.