ACE Robotics has open sourced its Kairos 3.0-4B embodied world model, enabling developers to build robots that can reason about the physical world and operate across multiple robot platforms with a single AI brain.
ACE Robotics has announced the open-source release of Kairos 3.0-4B, a native embodied intelligence world model designed to enable robots to understand and interact with the physical world. The model is publicly available on repositories including GitHub and Hugging Face, allowing developers and researchers to access and deploy it for robotics applications.
Described as the industry’s first native world model built specifically for embodied intelligence, Kairos 3.0-4B integrates multimodal understanding, generation, and prediction within a unified architecture. The model is designed from the ground up for real-world operation, incorporating physical laws, human behavioural patterns, and real robot action data to deliver physics-consistent reasoning.
The architecture embeds causal reasoning chains into its decision-making process, allowing robots not only to determine what actions to perform but also why those actions are appropriate. It integrates three core data sources—real robot interaction data, structured human behavioural data, and chain-of-thought reasoning data—to improve the reuse of real-world robotics datasets.
Despite its relatively compact size of four billion parameters and a memory requirement of 23.5GB VRAM, the model delivers strong performance. On A800 GPU benchmarks, Kairos 3.0-4B achieved inference speeds 72 times faster than NVIDIA Cosmos 2.5.
The model can also run in real time on the NVIDIA Jetson Thor T5000 edge platform, generating outputs 1.5 times faster than real-time video while issuing full-body commands for robot movement.
Kairos 3.0-4B supports cross-embodiment deployment across robots such as Agilex PIPER, Unitree G1, and Galaxy G1, enabling a single AI model to control multiple robotic forms and perform long-horizon interactions lasting up to seven minutes.
