© ROOT-NATION.com - Use of content is permitted with a backlink.
Today, during the GTC Taipei event at Computex 2026, Jensen Huang presented the world with Cosmos 3. This product is NVIDIA’s largest release in the field of open source software. The novelty is a basic model of physical AI that can combine visual analysis, environment modeling, and action prediction into one single complex.
At the same time, the American tech giant announced the creation of the Cosmos Coalition consortium, which includes Agile Robots, Black Forest Labs, Runway, and Skild AI. The joint efforts of these partners will be aimed at accelerating the development of publicly available models of the world around us. This announcement was the most significant news in the field of computer systems to date.
NVIDIA Cosmos 3 is positioned as the world’s first fully public omnimodel designed to implement physical AI and built on an innovative Transformers blend architecture. At a basic level, the Cosmos 3 software product is able to perceive and independently create textual content, graphic images, video materials, background sounds, and action algorithms within a single computing structure.
This opens up opportunities for the parallel generation of synthetic information and the design of physical AI behavioral strategies. Two versions of the Hugging Face platform are already available for developers: the more compact Cosmos 3 Nano and the full-size Cosmos 3 Super. All related tools, including scripts for training the neural network, tools for its deployment, and databases, are freely available on GitHub.
At the same time, NVIDIA launched the Cosmos Coalition, founded by Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI. The created Cosmos 3 model demonstrated the best performance indicators and topped the rankings in the tests for creating images from text and generating videos from static images. Among the brands that have already officially confirmed their intention to integrate this tool into their processes are Samsung, LG Electronics, Li Auto, and Doosan Robotics.
A large number of modern artificial intelligence systems are good at human language, and some developments are capable of analyzing graphic files and videos. However, physical AI technology faces a much more difficult challenge, which is to estimate future events in real space. For example, a robotic system’s manipulator must calculate the trajectory of its displacement after touching an object in advance. Similarly, a self-driving car must accurately analyze the likely behavior of a person who is about to step off the sidewalk onto the roadway. Such operations are not related to speech processing, as they lie in the realm of calculating the laws of physics.
The Cosmos 3 software package was created to solve such problems. The special architecture of the transformer mixture allows simultaneous processing of several streams of information coming from surveillance cameras, various sensors, or in the form of text instructions. Based on this information, the system predicts the future state of the environment and generates clear recommendations for further action. Thanks to this, the equipment running Cosmos 3 does not just record the current circumstances, but creates an internal simulation of what the space will look like in a second, and then makes decisions based on the forecast.
Previous versions of physical AI systems used separate isolated algorithms to process different types of data, where one part was responsible for vision, another for speech, and the third calculated direct movements. The combination of such components resembled a patchwork quilt. In contrast, Cosmos 3 works with text, pictures, video, ambient noise, and physical actions within a single balanced model. This architectural change is critical because in real-world environments, devices need to cross these boundaries instantly. They must simultaneously recognize voice commands, scan the environment, predict sound signals, and give orders to actuators within a single computation cycle. The Cosmos 3 Nano variant is a lightweight and less demanding version for use on endpoint devices, while the Cosmos 3 Super version is a full-scale data center solution focused on deep learning and high-fidelity process simulation.
NVIDIA Corporation has decided to make the digital weights of the model, its training scripts, service tools for implementation, and the full datasets on which Cosmos 3 was trained publicly available. This level of openness for advanced AI developments is a rare phenomenon. The direct goal of this step is to ensure reproducibility of results so that third-party research laboratories can independently verify NVIDIA’s achievements and create their own solutions based on this basis, instead of simply trusting the numbers in marketing materials. The Cosmos Coalition became an organizational tool for such cooperation. The members of this association, including Agile Robots, Black Forest Labs, and Runway, are committed to developing a joint ecosystem of open world models instead of developing closed competing products.
Read also:
- Huawei introduces Tau Scaling Law as a response to US chip sanctions: What is known
- New Zhenwu M890 AI chip: How U.S. sanctions contributed to China’s technological independence