A five-billion-parameter model just outperformed rivals nearly three times its size at predicting video in real time.
Astronex-World 1.0 is an open, controllable video world model that takes a text prompt or a starting image and predicts what happens next, frame by frame, while a user steers camera movement, continuous actions, and even inserts text events mid-rollout. It comes in two flavors: a bidirectional version for full-context generation, and a causal version built for real-time, streaming use, both based on the existing Wan2.2-TI2V-5B model. A camera-encoding technique called PRoPE feeds in intrinsics and extrinsics, and a 64-dimensional action stream runs through every layer of the transformer. Getting there took five training stages: teaching bidirectional camera and action control, converting the backbone to block-causal attention, distilling a fast few-step student model, restoring dynamics across mixed domains, and a final asymmetric DMD/DMD2 distribution-matching pass.
The causal model renders 832x480 video at 24 frames per second and streams in real time on a single Nvidia L20 GPU - training the whole pipeline took just two of those cards. On the WBench Full benchmark it scored 70.0, beating the 13.6-billion-parameter LongCat-Video and 14-billion-parameter Helios, landing within a point of the 22-billion-parameter LTX-2.3, and edging out YUME 1.5, a rival trained from the same 5B base model but on Nvidia A100 GPUs instead. That is a genuine efficiency result, not just a bigger-number one, and the reserved action interfaces mean it is built to be fine-tuned for robotics and self-driving work later.
Benchmark scores are not the same as a robot avoiding a wall, and open weights matched to real-time speed is exactly the combination every ambitious world-model launch promises but few actually deliver.