AMAP open-sources its world model DreamX-World

Alibaba's AMAP open-sources DreamX-World 1.0, an interactive 5B world model based on Wan 2.2 that generates up to one minute of controllable video.

AMAP, Alibaba's mapping subsidiary, open-sources DreamX-World 1.0, an interactive world model. Rather than simple passive video generation, the model produces environments that can be explored, controlled, and transformed via event prompts: scene movement, precise camera control, and triggered world modifications, in both first-person and third-person perspectives, across realistic or more stylized settings.

The released version, DreamX-World-5B, now generates up to one minute of autoregressive video, where an initial variant was limited to five seconds. It relies on Alibaba's Wan 2.2 video model and uses geometry-guided memory to preserve scene consistency when returning to it. Training combines data from Unreal Engine, game sequences, and real-world videos, followed by reinforcement learning and then distillation to accelerate inference.

The code, weights, and technical report are published under the Apache 2.0 license. An interactive beta version is also mentioned for the future, without a specified timeline.