AutoNavi Upgrades World Model ABot-World-0: Achieves the World's First 24-Hour Continuous Stable Reasoning

Recently, Alibaba's AutoNavi announced that its self-developed general interactive world model ABot-World-0 has completed a new round of technical upgrades. It now achieves continuous 24-hour stable inference on a single consumer-grade GPU, with no significant degradation in image quality, physical consistency, or interaction response. This makes it the first product in the world to extend the continuous inference duration of an interactive world model from "minutes" to "days."

For a long time, mainstream world models have generally maintained continuous generation capabilities of only 30 seconds to 1 minute. As the inference time increases, issues such as image breakdown, object deformation, distortion of physical laws, and misaligned interaction responses often occur. The root cause lies in the accumulation and gradual amplification of errors during the autoregressive video generation process.

To address this industry challenge, AutoNavi proposed a new long-range training paradigm called LongForcing, which unifies long-term stability and dynamic modeling of open scenarios into the training objective. During the training phase, the model continuously uses its own generated results as subsequent inputs, while constraints ensure that the generated content remains within a reasonable and stable world distribution, effectively alleviating error accumulation and distribution drift. This keeps image quality, scene structure, and dynamic evolution stable during ultra-long sequence inference.

Currently, ABot-World-0 is officially open-sourced, supporting local deployment on a single consumer-grade graphics card. Relevant capabilities have also been integrated into AutoNavi's previously released general world model development platform ABot-World Studio. This breakthrough further enhances the application potential of world models in scenarios such as long-term interaction, simulation, robot training, and digital twins, providing a new technical path for interactive world models to achieve long-term autonomous operation.