Astronex-World 1.0: Real-Time Interactive World Model Foundation
A 5B open video world model is being pitched as real-time, controllable, and competitive with much larger systems.
Astronex-World 1.0 takes text or an initial image and rolls forward video under camera paths, continuous actions, embodiment IDs, and inserted text events. The report describes both bidirectional generation and a causal streaming model built on Wan2.2-TI2V-5B, with the causal version producing 832x480 video at 24 fps. The authors say training runs on two NVIDIA L20 48 GB GPUs and real-time streaming runs on one. They report WBench scores that put it above several larger named models, while leaving action interfaces reserved for later embodied-intelligence and autonomous-driving post-training. ArXiv · AI/CL/LG's note
Astronex-World 1.0 takes text or an initial image and rolls forward video under camera paths, continuous actions, embodiment IDs, and inserted text events. The report describes both bidirectional generation and a causal streaming model built on Wan2.2-TI2V-5B, with the causal version producing 832x480 video at 24 fps. The authors say training runs on two NVIDIA L20 48 GB GPUs and real-time streaming runs on one. They report WBench scores that put it above several larger named models, while leaving action interfaces reserved for later embodied-intelligence and autonomous-driving post-training. ArXiv · AI/CL/LG's note
score 6